HB (@herbcase7) header
HB (@herbcase7) profile picture

SPiFS Team Leader / Honeyinfo founder https://teapawt.org/t.co/Beu66LyeAd / https://teapawt.org/t.co/3aXkqGRO6z Ambassador / FDS pm mng. / KOL / lp provider / KR,JP,EN

Seoul, Korea Rep.
Joined March 2021
Photo by HB (@herbcase7)Photo by HB (@herbcase7)Photo by HB (@herbcase7)Photo by HB (@herbcase7)Photo by HB (@herbcase7)Photo by HB (@herbcase7)Photo by HB (@herbcase7)Photo by HB (@herbcase7)Photo by HB (@herbcase7)Photo by HB (@herbcase7)
๐Ÿ“š Rialo Research Series 67: Agent Memory Poisoning - Can an Agent's Past Be Weaponized? (Kor : ์—์ด์ „ํŠธ ๋ฉ”๋ชจ๋ฆฌ ์˜ค์—ผ - AI Agent์˜ ๊ณผ๊ฑฐ ๊ธฐ์–ต์ด ๊ณต๊ฒฉ ๋„๊ตฌ๊ฐ€ ๋  ์ˆ˜ ์žˆ์„๊นŒ?) 1/ Intro: When Memory Becomes an Attack Surface (EN) In the previous chapter, we explored why an autonomous agent's internal state must be verified before execution resumes. But what if the agent's memory is deliberately manipulated without triggering an obvious system failure? Imagine a lending agent that remembers historical market conditions, borrower behavior, and previous risk assessments. An attacker inserts misleading information into a source the agent regularly consults. The agent continues operating normally, unaware that its future decisions are being influenced. The most dangerous attack may not interrupt execution. It may quietly change what the agent believes. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” Autonomous Agent์˜ ์‹คํ–‰ ์žฌ๊ฐœ ์ „์— Internal State๋ฅผ ๊ฒ€์ฆํ•ด์•ผ ํ•˜๋Š” ์ด์œ ๋ฅผ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์‹œ์Šคํ…œ ์žฅ์•  ์—†์ด Agent์˜ Memory๊ฐ€ ์˜๋„์ ์œผ๋กœ ์กฐ์ž‘๋œ๋‹ค๋ฉด ์–ด๋–จ๊นŒ์š”? ๊ณผ๊ฑฐ ์‹œ์žฅ ์ƒํ™ฉ, Borrower Behavior, Risk Assessment๋ฅผ ๊ธฐ์–ตํ•˜๋Š” Lending Agent๋ฅผ ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์ž๊ฐ€ Agent๊ฐ€ ์ •๊ธฐ์ ์œผ๋กœ ์ฐธ๊ณ ํ•˜๋Š” ๋ฐ์ดํ„ฐ์— ์ž˜๋ชป๋œ ์ •๋ณด๋ฅผ ์‚ฝ์ž…ํ•ฉ๋‹ˆ๋‹ค. Agent๋Š” ์ •์ƒ์ ์œผ๋กœ ์ž‘๋™ํ•˜์ง€๋งŒ ์ดํ›„ ํŒ๋‹จ์€ ์ด๋ฏธ ์˜ค์—ผ๋œ ์ •๋ณด์˜ ์˜ํ–ฅ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค. ๊ฐ€์žฅ ์œ„ํ—˜ํ•œ ๊ณต๊ฒฉ์€ ์‹คํ–‰์„ ์ค‘๋‹จ์‹œํ‚ค๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ Agent๊ฐ€ ๋ฏฟ๋Š” ์ •๋ณด๋ฅผ ์กฐ์šฉํžˆ ๋ฐ”๊พธ๋Š” ๊ฒƒ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 2/ How Memory Poisoning Works (EN) An agent may collect information from websites, APIs, documents, or other agents. Some of that information can eventually become part of its persistent memory. An attacker could exploit this process by introducing fabricated facts, misleading summaries, or instructions disguised as trusted records. For example, a malicious document might claim that a particular collateral asset has received special risk approval. If the agent stores that claim without verification, future lending decisions could rely on a false assumption. (KR) Agent๋Š” Website, API, Document, ๋‹ค๋ฅธ Agent ๋“ฑ ๋‹ค์–‘ํ•œ ์ถœ์ฒ˜์—์„œ ์ •๋ณด๋ฅผ ์ˆ˜์ง‘ํ•ฉ๋‹ˆ๋‹ค. ์ด ์ค‘ ์ผ๋ถ€๋Š” Persistent Memory์— ์ €์žฅ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์ž๋Š” ์ด ๊ณผ์ •์„ ์•…์šฉํ•ด ์กฐ์ž‘๋œ ์‚ฌ์‹ค์ด๋‚˜ ์ž˜๋ชป๋œ Summary, ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋ก์ฒ˜๋Ÿผ ์œ„์žฅํ•œ Instruction์„ ์ฃผ์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์•…์˜์ ์ธ ๋ฌธ์„œ๊ฐ€ ํŠน์ • ๋‹ด๋ณด ์ž์‚ฐ์— ํŠน๋ณ„ํ•œ Risk Approval์ด ๋ถ€์—ฌ๋๋‹ค๊ณ  ์ฃผ์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Agent๊ฐ€ ์ด๋ฅผ ๊ฒ€์ฆ ์—†์ด ์ €์žฅํ•œ๋‹ค๋ฉด ์ดํ›„ Lending Decision์€ ์ž˜๋ชป๋œ ์ „์ œ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ด๋ฃจ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 3/ Why Delayed Attacks Are Difficult to Detect (EN) Memory poisoning does not always produce immediate damage. An attacker might introduce false information today, while the affected decision occurs weeks later. By then, the original source may no longer appear in the agent's active context. The agent simply retrieves a stored memory and treats it as established knowledge. This delay makes it difficult to connect the eventual financial loss to the original manipulation. A successful attack can survive long after the malicious input disappears. (KR) Memory Poisoning์ด ๋ฐ˜๋“œ์‹œ ์ฆ‰๊ฐ์ ์ธ ํ”ผํ•ด๋ฅผ ๋ฐœ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์ž๊ฐ€ ์˜ค๋Š˜ ์ž˜๋ชป๋œ ์ •๋ณด๋ฅผ ์ฃผ์ž…ํ–ˆ๋”๋ผ๋„ ์‹ค์ œ ๊ธˆ์œต ํ”ผํ•ด๋Š” ๋ช‡ ์ฃผ ๋’ค์— ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋•Œ๋Š” ์ตœ์ดˆ ์ •๋ณด ์ถœ์ฒ˜๊ฐ€ Agent์˜ Active Context์— ๋‚จ์•„ ์žˆ์ง€ ์•Š์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. Agent๋Š” ์ €์žฅ๋œ Memory๋ฅผ ๋ถˆ๋Ÿฌ์™€ ์ด๋ฏธ ๊ฒ€์ฆ๋œ ์ง€์‹์ฒ˜๋Ÿผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์‹œ๊ฐ„์ฐจ ๋•Œ๋ฌธ์— ์‹ค์ œ ๊ธˆ์œต ์†์‹ค๊ณผ ์ตœ์ดˆ ๊ณต๊ฒฉ์„ ์—ฐ๊ฒฐํ•˜๊ธฐ ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. ์•…์˜์ ์ธ Input์ด ์‚ฌ๋ผ์ง„ ์ดํ›„์—๋„ ๊ณต๊ฒฉ์˜ ์˜ํ–ฅ์€ ๊ณ„์† ๋‚จ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 4/ Memory Provenance and Trust Boundaries (EN) One possible defense is recording where each important memory originated. A memory entry could include its source, timestamp, verification status, and the permissions under which it was created. Information retrieved from an external website should not automatically receive the same trust level as an authenticated protocol state. Applications could also separate unverified observations from approved operational rules. This helps prevent external content from silently becoming execution authority. (KR) ํ•œ ๊ฐ€์ง€ ๋ฐฉ์–ด ๋ฐฉ๋ฒ•์€ ์ค‘์š”ํ•œ Memory๊ฐ€ ์–ด๋””์—์„œ ์ƒ์„ฑ๋๋Š”์ง€ ๊ธฐ๋กํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. Memory Entry์— Source, Timestamp, Verification Status, ์ƒ์„ฑ ๋‹น์‹œ Permission ๋“ฑ์„ ์—ฐ๊ฒฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์™ธ๋ถ€ Website์—์„œ ์ˆ˜์ง‘ํ•œ ์ •๋ณด๊ฐ€ ์ธ์ฆ๋œ Protocol State์™€ ๋™์ผํ•œ ์‹ ๋ขฐ ์ˆ˜์ค€์„ ๊ฐ€์ ธ์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค. ๊ฒ€์ฆ๋˜์ง€ ์•Š์€ Observation๊ณผ ์Šน์ธ๋œ Operational Rule์„ ๋ถ„๋ฆฌํ•˜๋Š” ๋ฐฉ๋ฒ•๋„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์™ธ๋ถ€์—์„œ ์œ ์ž…๋œ ์ •๋ณด๊ฐ€ ๊ฒ€์ฆ ์—†์ด ์‹คํ–‰ ๊ถŒํ•œ์œผ๋กœ ์—ฐ๊ฒฐ๋˜์ง€ ์•Š๋„๋ก ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. 5/ The Limits of Simple Memory Cleaning (EN) Deleting a suspicious memory entry may not fully eliminate its influence. The agent could have generated additional summaries, risk scores, or execution plans based on that information. Other agents may also have received and stored the contaminated conclusions. Recovery therefore requires identifying downstream dependencies, not merely removing the original entry. Affected calculations may need to be repeated using verified inputs. Cleaning one memory does not guarantee that every decision derived from it has been corrected. (KR) ์˜์‹ฌ์Šค๋Ÿฌ์šด Memory Entry๋ฅผ ์‚ญ์ œํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ ์˜ํ–ฅ์ด ์™„์ „ํžˆ ์‚ฌ๋ผ์ง€์ง€๋Š” ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Agent๊ฐ€ ์ด๋ฏธ ํ•ด๋‹น ์ •๋ณด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Summary, Risk Score, Execution Plan ๋“ฑ์„ ์ƒ์„ฑํ–ˆ์„ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋‹ค๋ฅธ Agent์—๊ฒŒ ์˜ค์—ผ๋œ ํŒ๋‹จ ๊ฒฐ๊ณผ๊ฐ€ ์ „๋‹ฌ๋์„ ๊ฐ€๋Šฅ์„ฑ๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ตœ์ดˆ Memory๋ฅผ ์‚ญ์ œํ•˜๋Š” ๊ฒƒ๋ฟ ์•„๋‹ˆ๋ผ ์ดํ›„ ์ƒ์„ฑ๋œ ์ •๋ณด์˜ ์˜์กด ๊ด€๊ณ„๊นŒ์ง€ ํ™•์ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์˜ํ–ฅ์„ ๋ฐ›์€ ๊ณ„์‚ฐ์€ ๊ฒ€์ฆ๋œ Input์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋‹ค์‹œ ์ˆ˜ํ–‰ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜๋‚˜์˜ ๊ธฐ์–ต์„ ์‚ญ์ œํ–ˆ๋‹ค๊ณ  ๊ทธ ๊ธฐ์–ต์—์„œ ํŒŒ์ƒ๋œ ๋ชจ๋“  ํŒ๋‹จ๊นŒ์ง€ ์ˆ˜์ •๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. 6/ Rialo and Memory Integrity Controls (EN) Rialo's Reactive Transactions provide an interesting foundation for designing responses to detected integrity violations. Applications could use predefined conditions to restrict sensitive transactions when a critical memory reference fails verification. Execution records could also help connect financial actions to the application state and authorization conditions used at execution time. However, reactive execution does not automatically detect poisoned AI memory or verify the truthfulness of external information. Memory provenance, validation policies, and dependency tracking would still require application-level implementation. (KR) Rialo์˜ Reactive Transactions๋Š” Memory Integrity ์œ„๋ฐ˜์ด ๊ฐ์ง€๋์„ ๋•Œ ๋Œ€์‘ํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ค‘์š”ํ•œ Memory Reference์˜ ๊ฒ€์ฆ์ด ์‹คํŒจํ•˜๋ฉด ์‚ฌ์ „์— ์ •์˜๋œ ์กฐ๊ฑด์— ๋”ฐ๋ผ ๊ณ ์œ„ํ—˜ Transaction์„ ์ œํ•œํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. Execution Record ์—ญ์‹œ ๊ธˆ์œต ํ–‰๋™๊ณผ ์‹คํ–‰ ๋‹น์‹œ Application State, Authorization Condition์„ ์—ฐ๊ฒฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ Reactive Execution์ด ์˜ค์—ผ๋œ AI Memory๋ฅผ ์ž๋™์œผ๋กœ ํƒ์ง€ํ•˜๊ฑฐ๋‚˜ ์™ธ๋ถ€ ์ •๋ณด์˜ ์ง„์œ„๋ฅผ ๊ฒ€์ฆํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. Memory Provenance, Validation Policy, Dependency Tracking ๋“ฑ์€ Application ์ฐจ์›์—์„œ ๊ตฌํ˜„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a demonstration where a lending agent receives a malicious document containing fabricated collateral information. The agent initially stores the information as an unverified observation. When a later transaction attempts to use that memory for a risk decision, the application checks its provenance. Verification fails, and the transaction is restricted. The system then identifies related summaries and recalculates affected risk assessments using authenticated data. This would demonstrate how memory verification can prevent delayed attacks from becoming financial losses. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์กฐ์ž‘๋œ ๋‹ด๋ณด ์ •๋ณด๊ฐ€ ํฌํ•จ๋œ ์•…์˜์ ์ธ Document๋ฅผ Lending Agent๊ฐ€ ์ˆ˜์‹ ํ•˜๋Š” Demo์ž…๋‹ˆ๋‹ค. Agent๋Š” ํ•ด๋‹น ์ •๋ณด๋ฅผ ์ฒ˜์Œ์—๋Š” Unverified Observation์œผ๋กœ ์ €์žฅํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„ Transaction์ด ํ•ด๋‹น Memory๋ฅผ Risk Decision์— ํ™œ์šฉํ•˜๋ ค๊ณ  ํ•˜๋ฉด Application์ด Provenance๋ฅผ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ฆ์— ์‹คํŒจํ•˜๋ฉด Transaction ์‹คํ–‰์ด ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ์ด์–ด์„œ ๊ด€๋ จ Summary๋ฅผ ์ถ”์ ํ•˜๊ณ  ์ธ์ฆ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Risk Assessment๋ฅผ ๋‹ค์‹œ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค. Memory Verification์ด ์‹œ๊ฐ„์ฐจ๋ฅผ ๋‘๊ณ  ๋ฐœ์ƒํ•˜๋Š” ๊ณต๊ฒฉ์„ ์‹ค์ œ ๊ธˆ์œต ํ”ผํ•ด๋กœ ์ด์–ด์ง€๊ธฐ ์ „์— ์ฐจ๋‹จํ•˜๋Š” ๊ณผ์ •์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. (EN) As autonomous agents become more dependent on persistent memory, their security boundaries must extend beyond wallets, transactions, and execution permissions. An agent can follow every execution rule correctly while making decisions based on information that was deliberately corrupted. Protecting autonomous finance therefore requires understanding not only what an agent does, but where its knowledge comes from and whether that knowledge remains trustworthy. An agent's memory should be treated as evidence to verify, not unquestionable truth to execute. (KR) Autonomous Agent๊ฐ€ Persistent Memory์— ๋” ๋งŽ์ด ์˜์กดํ• ์ˆ˜๋ก ๋ณด์•ˆ ๊ฒฝ๊ณ„๋„ Wallet, Transaction, Execution Permission์„ ๋„˜์–ด ํ™•์žฅ๋ผ์•ผ ํ•ฉ๋‹ˆ๋‹ค. Agent๋Š” ๋ชจ๋“  ์‹คํ–‰ ๊ทœ์น™์„ ์ •ํ™•ํ•˜๊ฒŒ ์ค€์ˆ˜ํ•˜๋ฉด์„œ๋„ ์˜๋„์ ์œผ๋กœ ์กฐ์ž‘๋œ ์ •๋ณด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ž˜๋ชป๋œ ๊ฒฐ์ •์„ ๋‚ด๋ฆด ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Autonomous Finance์˜ ๋ณด์•ˆ์„ ์œ„ํ•ด์„œ๋Š” Agent๊ฐ€ ๋ฌด์—‡์„ ์‹คํ–‰ํ•˜๋Š”์ง€๋งŒ ํ™•์ธํ•ด์„œ๋Š” ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. Agent๊ฐ€ ์‚ฌ์šฉํ•˜๋Š” ์ง€์‹์ด ์–ด๋””์—์„œ ์™”๊ณ , ์ง€๊ธˆ๋„ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ Agent์˜ Memory๋Š” ๋ฌด์กฐ๊ฑด ์‹ ๋ขฐํ•ด์•ผ ํ•  ์ •๋‹ต์ด ์•„๋‹ˆ๋ผ, ์ถœ์ฒ˜์™€ ๋ฌด๊ฒฐ์„ฑ์„ ์ง€์†์ ์œผ๋กœ ๊ฒ€์ฆํ•ด์•ผ ํ•˜๋Š” ํŒ๋‹จ ๊ทผ๊ฑฐ๊ฐ€ ์•„๋‹๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
๐Ÿ“š Rialo Research Series 66: Agent State Integrity - Can a Recovered Agent Trust Its Own Memory? (Kor : ์—์ด์ „ํŠธ ์ƒํƒœ ๋ฌด๊ฒฐ์„ฑ - ๋ณต๊ตฌ๋œ Agent๋Š” ์ž์‹ ์˜ ๊ธฐ์–ต์„ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ์„๊นŒ?) 1/ Intro: Recovery Doesn't Guarantee Integrity (EN) In the previous chapter, we explored how a suspended autonomous agent could safely regain execution authority. But restoring permissions doesn't necessarily mean the agent's internal state is trustworthy. Imagine a lending agent suspended after receiving manipulated oracle prices. The data source is corrected, its permissions are reviewed, and execution resumes. However, the agent still remembers an incorrect liquidation threshold calculated before the shutdown. An agent can be operationally recovered while remaining logically compromised. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ์ค‘๋‹จ๋œ Autonomous Agent๊ฐ€ ์–ด๋–ป๊ฒŒ ์•ˆ์ „ํ•˜๊ฒŒ Execution Authority๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ๊ถŒํ•œ์„ ๋ณต๊ตฌํ–ˆ๋‹ค๊ณ  ํ•ด์„œ Agent์˜ Internal State๊นŒ์ง€ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์กฐ์ž‘๋œ Oracle ๊ฐ€๊ฒฉ์„ ๋ฐ›์•„ ์ค‘๋‹จ๋œ Lending Agent๋ฅผ ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๊ฐ€ ์ˆ˜์ •๋˜๊ณ  Permission Review๊นŒ์ง€ ์™„๋ฃŒ๋ผ ์‹คํ–‰์„ ์žฌ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Agent ๋‚ด๋ถ€์—๋Š” ์ค‘๋‹จ ์ด์ „์— ๊ณ„์‚ฐํ•œ ์ž˜๋ชป๋œ Liquidation Threshold๊ฐ€ ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. Agent๋Š” ์ •์ƒ์ ์œผ๋กœ ์žฌ๊ฐ€๋™๋๋”๋ผ๋„ ๋‚ด๋ถ€ ํŒ๋‹จ ์ƒํƒœ๋Š” ์—ฌ์ „ํžˆ ์˜ค์—ผ๋ผ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 2/ What Makes Up an Agent's State? (EN) An autonomous agent's state may contain more than wallet balances or transaction history. It can include cached market data, risk calculations, execution queues, temporary permissions, and previous decisions. Some information may be stored on-chain, while other components remain in off-chain databases or agent memory. Verifying only blockchain transactions therefore cannot establish the integrity of the entire agent. A reliable recovery process needs to identify which state components influenced the failed decision. (KR) Autonomous Agent์˜ State๋Š” Wallet Balance๋‚˜ Transaction History๋งŒ์œผ๋กœ ๊ตฌ์„ฑ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. Cached Market Data, Risk Calculation, Execution Queue, Temporary Permission, Previous Decision ๋“ฑ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ์ •๋ณด๋Š” On-chain์— ๊ธฐ๋ก๋˜์ง€๋งŒ ๋‹ค๋ฅธ ์ •๋ณด๋Š” Off-chain Database๋‚˜ Agent Memory์— ๋‚จ์•„ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Blockchain Transaction๋งŒ ๊ฒ€์ฆํ•œ๋‹ค๊ณ  Agent ์ „์ฒด์˜ ๋ฌด๊ฒฐ์„ฑ์ด ๋ณด์žฅ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์‹คํŒจํ•œ ํŒ๋‹จ์— ์–ด๋–ค State Component๊ฐ€ ์˜ํ–ฅ์„ ๋ฏธ์ณค๋Š”์ง€ ํ™•์ธํ•˜๋Š” ๊ณผ์ •์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 3/ The Danger of Contaminated Memory (EN) Suppose an agent stores a summary stating that a particular asset has become dangerously undercollateralized. That summary was generated from manipulated prices. Even after the oracle is corrected, the agent might continue using the stored conclusion instead of recalculating the risk. This becomes especially problematic when agents rely on persistent memory to make decisions across multiple sessions. The original data may be fixed, but the incorrect interpretation can survive. (KR) Agent๊ฐ€ ํŠน์ • ์ž์‚ฐ์˜ ๋‹ด๋ณด ๋น„์œจ์ด ์œ„ํ—˜ ์ˆ˜์ค€์œผ๋กœ ๋–จ์–ด์กŒ๋‹ค๋Š” Summary๋ฅผ ์ €์žฅํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ํ•ด๋‹น Summary๋Š” ์กฐ์ž‘๋œ ๊ฐ€๊ฒฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋์Šต๋‹ˆ๋‹ค. Oracle์ด ์ˆ˜์ •๋œ ์ดํ›„์—๋„ Agent๊ฐ€ Risk๋ฅผ ๋‹ค์‹œ ๊ณ„์‚ฐํ•˜์ง€ ์•Š๊ณ  ๊ธฐ์กด ๊ฒฐ๋ก ์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์—ฌ๋Ÿฌ Session์— ๊ฑธ์ณ Persistent Memory๋ฅผ ์‚ฌ์šฉํ•˜๋Š” Agent๋ผ๋ฉด ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ๋ณต์žกํ•ด์ง‘๋‹ˆ๋‹ค. ์›๋ณธ ๋ฐ์ดํ„ฐ๋Š” ์ˆ˜์ •๋์ง€๋งŒ ์ž˜๋ชป๋œ ํ•ด์„์€ Memory์— ๊ณ„์† ๋‚จ์•„ ์žˆ์„ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. 4/ State Reconstruction and Verification (EN) One possible approach is reconstructing critical state from verified inputs. The application could identify the last trusted checkpoint, retrieve authenticated data, and recompute affected risk assessments. Stored decisions would then be compared with newly calculated results. Any mismatch could trigger further investigation before execution resumes. However, reconstructing state requires access to reliable historical inputs and deterministic processing where applicable. A checkpoint is only useful if its own integrity can be established. (KR) ํ•œ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์€ ๊ฒ€์ฆ๋œ Input์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ค‘์š”ํ•œ State๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. Application์€ ๋งˆ์ง€๋ง‰์œผ๋กœ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” Checkpoint๋ฅผ ํ™•์ธํ•˜๊ณ , ์ธ์ฆ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ถˆ๋Ÿฌ์™€ ์˜ํ–ฅ์„ ๋ฐ›์€ Risk Assessment๋ฅผ ๋‹ค์‹œ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ์ €์žฅ๋œ Decision๊ณผ ์ƒˆ๋กญ๊ฒŒ ๊ณ„์‚ฐํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋ถˆ์ผ์น˜๊ฐ€ ๋ฐœ๊ฒฌ๋˜๋ฉด ์‹คํ–‰ ์žฌ๊ฐœ ์ „์— ์ถ”๊ฐ€ ๊ฒ€์ฆ์„ ์ง„ํ–‰ํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๋‹ค๋งŒ State Reconstruction์—๋Š” ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” Historical Input๊ณผ ํ•„์š”ํ•œ ๊ฒฝ์šฐ ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ๊ณ„์‚ฐ ๊ณผ์ •์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. Checkpoint ์—ญ์‹œ ๋ฌด๊ฒฐ์„ฑ์ด ํ™•์ธ๋ผ์•ผ ๋ณต๊ตฌ ๊ธฐ์ค€์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 5/ When State Cannot Be Fully Reconstructed (EN) Not every agent decision can be perfectly reproduced. An AI model may generate different outputs when its model version, context, or external dependencies change. Some off-chain information may also be unavailable after an incident. In these situations, pretending that a complete reconstruction is possible would create false confidence. Applications may need to invalidate uncertain state, restrict sensitive actions, and require additional authorization before restoring execution. Unverifiable state should not automatically become trusted state. (KR) ๋ชจ๋“  Agent์˜ ํŒ๋‹จ์„ ์™„๋ฒฝํ•˜๊ฒŒ ์žฌํ˜„ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. AI Model Version์ด๋‚˜ Context, External Dependency๊ฐ€ ๋ณ€๊ฒฝ๋˜๋ฉด ๊ฐ™์€ ์ž…๋ ฅ์—์„œ๋„ ๋‹ค๋ฅธ ๊ฒฐ๊ณผ๊ฐ€ ์ƒ์„ฑ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ Off-chain Information์€ ์‚ฌ๊ณ  ์ดํ›„ ํ™•๋ณดํ•˜์ง€ ๋ชปํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฐ ์ƒํ™ฉ์—์„œ ์™„์ „ํ•œ Reconstruction์ด ๊ฐ€๋Šฅํ•˜๋‹ค๊ณ  ๊ฐ€์ •ํ•˜๋ฉด ์˜คํžˆ๋ ค ์ž˜๋ชป๋œ ์‹ ๋ขฐ๊ฐ€ ํ˜•์„ฑ๋ฉ๋‹ˆ๋‹ค. Application์€ ๊ฒ€์ฆํ•  ์ˆ˜ ์—†๋Š” State๋ฅผ ๋ฌดํšจํ™”ํ•˜๊ฑฐ๋‚˜ ๊ณ ์œ„ํ—˜ ํ–‰๋™์„ ์ œํ•œํ•˜๊ณ , ์‹คํ–‰ ๊ถŒํ•œ ๋ณต๊ตฌ์— ์ถ”๊ฐ€ ์Šน์ธ์„ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฆํ•  ์ˆ˜ ์—†๋Š” ์ƒํƒœ๋ฅผ ๋‹จ์ˆœํžˆ ์ •์ƒ ์ƒํƒœ๋กœ ๊ฐ„์ฃผํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค. 6/ Rialo and State Integrity (EN) Rialo's execution-oriented architecture offers an interesting environment for exploring state integrity checks. Reactive Transactions could help applications trigger verification workflows when predefined inconsistencies are detected. Execution records and application-managed checkpoints could support investigations into how an agent's state changed. However, Rialo's reactive capabilities do not automatically verify arbitrary off-chain memory or guarantee deterministic AI outputs. Developers must define which state elements require verification, how trusted references are established, and what happens when validation fails. (KR) Rialo์˜ Execution ์ค‘์‹ฌ ์•„ํ‚คํ…์ฒ˜๋Š” State Integrity ๊ฒ€์ฆ์„ ํƒ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ํ™˜๊ฒฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Reactive Transactions๋ฅผ ํ™œ์šฉํ•˜๋ฉด ์‚ฌ์ „์— ์ •์˜๋œ ๋ถˆ์ผ์น˜๊ฐ€ ๊ฐ์ง€๋์„ ๋•Œ Verification Workflow๋ฅผ ์‹œ์ž‘ํ•˜๋„๋ก ์„ค๊ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Execution Record์™€ Application์ด ๊ด€๋ฆฌํ•˜๋Š” Checkpoint๋Š” Agent State๊ฐ€ ์–ด๋–ป๊ฒŒ ๋ณ€๊ฒฝ๋๋Š”์ง€ ์กฐ์‚ฌํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ Rialo์˜ Reactive ๊ธฐ๋Šฅ์ด ๋ชจ๋“  Off-chain Memory๋ฅผ ์ž๋™์œผ๋กœ ๊ฒ€์ฆํ•˜๊ฑฐ๋‚˜ AI Output์˜ ๊ฒฐ์ •์„ฑ์„ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์–ด๋–ค State๋ฅผ ๊ฒ€์ฆํ•˜๊ณ , ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ์ค€์„ ์–ด๋–ป๊ฒŒ ์„ค์ •ํ•˜๋ฉฐ, ๊ฒ€์ฆ ์‹คํŒจ ์‹œ ์–ด๋–ค ์กฐ์น˜๋ฅผ ์ทจํ• ์ง€๋Š” ๊ฐœ๋ฐœ์ž๊ฐ€ ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a demonstration where a lending agent is suspended after using manipulated oracle data. The oracle is corrected, and the agent requests permission to resume. Before approval, the application compares the agent's stored risk assessment against a newly reconstructed state. The verification process discovers that the agent's memory still contains an incorrect liquidation threshold. Execution remains restricted until the contaminated state is invalidated and the calculation is repeated. This would demonstrate why restoring execution authority and restoring trustworthy decision-making are separate requirements. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์กฐ์ž‘๋œ Oracle ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ ๋’ค ์ค‘๋‹จ๋œ Lending Agent์˜ ๋ณต๊ตฌ Demo์ž…๋‹ˆ๋‹ค. Oracle์ด ์ •์ƒํ™”๋˜๊ณ  Agent๊ฐ€ ์‹คํ–‰ ์žฌ๊ฐœ๋ฅผ ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์Šน์ธ ์ „์— Application์ด ์ €์žฅ๋œ Risk Assessment์™€ ์ƒˆ๋กญ๊ฒŒ ์žฌ๊ตฌ์„ฑํ•œ State๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ฆ ๊ณผ์ •์—์„œ Agent Memory์— ์ž˜๋ชป๋œ Liquidation Threshold๊ฐ€ ๋‚จ์•„ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐœ๊ฒฌ๋ฉ๋‹ˆ๋‹ค. ์˜ค์—ผ๋œ State๋ฅผ ๋ฌดํšจํ™”ํ•˜๊ณ  ๊ณ„์‚ฐ์„ ๋‹ค์‹œ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์ „๊นŒ์ง€ Execution์€ ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ์‹คํ–‰ ๊ถŒํ•œ ๋ณต๊ตฌ์™€ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ํŒ๋‹จ ๋Šฅ๋ ฅ์˜ ๋ณต๊ตฌ๊ฐ€ ์„œ๋กœ ๋‹ค๋ฅธ ์กฐ๊ฑด์ด๋ผ๋Š” ์ ์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. (EN) Autonomous finance increasingly depends on agents that remember previous decisions and operate continuously across multiple environments. Persistent memory improves efficiency, but it can also preserve mistakes long after their original causes disappear. A trustworthy recovery process must therefore examine not only what an agent is allowed to do, but also what information it still believes. An autonomous agent should not regain financial authority simply because it can run again. Its decision state must be trustworthy enough to justify that authority. (KR) Autonomous Finance์—์„œ๋Š” ์ด์ „ ํŒ๋‹จ์„ ๊ธฐ์–ตํ•˜๊ณ  ์—ฌ๋Ÿฌ ํ™˜๊ฒฝ์—์„œ ์ง€์†์ ์œผ๋กœ ์ž‘๋™ํ•˜๋Š” Agent์˜ ์ค‘์š”์„ฑ์ด ์ปค์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Persistent Memory๋Š” ํšจ์œจ์„ฑ์„ ๋†’์ด์ง€๋งŒ, ์ตœ์ดˆ ์˜ค๋ฅ˜๊ฐ€ ์‚ฌ๋ผ์ง„ ์ดํ›„์—๋„ ์ž˜๋ชป๋œ ํŒ๋‹จ์„ ๋ณด์กดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” Recovery Process๋ผ๋ฉด Agent๊ฐ€ ๋ฌด์—‡์„ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋งŒ ํ™•์ธํ•ด์„œ๋Š” ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. Agent๊ฐ€ ์–ด๋–ค ์ •๋ณด๋ฅผ ์‚ฌ์‹ค๋กœ ์ธ์‹ํ•˜๊ณ  ์žˆ๋Š”์ง€๊นŒ์ง€ ๊ฒ€์ฆํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ Autonomous Finance์—์„œ ์ค‘์š”ํ•œ ๊ฒƒ์€ Agent๋ฅผ ๋‹ค์‹œ ์ž‘๋™์‹œํ‚ค๋Š” ๊ฒƒ๋ฟ ์•„๋‹ˆ๋ผ, ๊ทธ Agent๊ฐ€ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ƒํƒœ์—์„œ ๋‹ค์‹œ ํŒ๋‹จํ•˜๋„๋ก ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ์•„๋‹๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
1
2
84
GTA6 ์ถœ์‹œ๋ฅผ ๋…ธ๋ฆฐ ๊ฐ€์งœ ์‚ฌ์ „ ์˜ˆ์•ฝ๊ณผ ์—์–ด๋“œ๋ž, ๋ธ”๋ก์—์ด๋“œ ๋ณด๊ณ ์„œ๋กœ ๋ณธ ์ง€๊ฐ‘ ํƒˆ์ทจ ์ˆ˜๋ฒ• GTA6๊ฐ€ ํ•œ๊ตญ ์‹œ๊ฐ„ 11์›” 19์ผ 0์‹œ์— ์ •์‹ ์ถœ์‹œ๋๋Š”๋ฐ ์ถœ์‹œ ์•ฝ 3์‹œ๊ฐ„30๋ถ„ ๋’ค์ธ ์ƒˆ๋ฒฝ 3์‹œ30๋ถ„์ฏค ๋ธ”๋ก์ฒด์ธ ๋ณด์•ˆ์—…์ฒด ๋ธ”๋ก์—์ด๋“œ๊ฐ€ GTA6 ์ถœ์‹œ ์—ด๊ธฐ๋ฅผ ๋…ธ๋ฆฐ ํฌ๋ฆฝํ†  ์‚ฌ๊ธฐ ๋ณด๊ณ ์„œ๋ฅผ ๊ณต๊ฐœํ–ˆ๋‹ค. ๊ฐ€์งœ ์‚ฌ์ „ ์˜ˆ์•ฝ ํŽ˜์ด์ง€์™€ ์—์–ด๋“œ๋ž ํŽ˜์ด์ง€๊ฐ€ ๊ฒฐ์ œ ๋Œ€์‹  ์ง€๊ฐ‘ ์—ฐ๊ฒฐ์„ ์š”๊ตฌํ•˜๊ณ  ๊ทธ ์—ฐ๊ฒฐ ๋‹จ๊ณ„์—์„œ ์ž์‚ฐ์„ ๋นผ ๊ฐ€๋Š” ๊ตฌ์กฐ๋‹ค. ๋ธ”๋ก์—์ด๋“œ๋Š” 10๋…„ ๋„˜๊ฒŒ ๊ธฐ๋‹ค๋ฆฐ ๊ฒŒ์ž„์ด๋ผ๋Š” ์ ์„ ์‚ฌ๊ธฐ๊พผ๋“ค์ด ๋ชฐ๋ฆฐ ์ด์œ ๋กœ ๊ผฝ์•˜๋‹ค. ๊ฒŒ์ž„์„ ์‚ฌ๋ ค๋‹ค ์ง€๊ฐ‘์„ ์—ฐ๊ฒฐํ•˜๋Š” ์ˆœ๊ฐ„ ํ„ธ๋ฆฌ๋Š” ์ˆ˜๋ฒ•์ด๋ผ์„œ ํฌ๋ฆฝํ†  ์ง€๊ฐ‘์„ ์“ฐ๋Š” ๊ฒŒ์ด๋จธ๋ผ๋ฉด ํ•œ ๋ฒˆ ์ •๋ฆฌํ•ด ๋‘˜ ๋งŒํ•˜๋‹ค. ์‚ฌ๊ธฐ๋Š” ์–ด๋–ป๊ฒŒ ์ง„ํ–‰๋ ๊นŒ? ๋ธ”๋ก์—์ด๋“œ๊ฐ€ ์„ค๋ช…ํ•˜๋Š” ํ๋ฆ„์€ ๋‹จ์ˆœํ•œ๋ฐ ์†Œ์…œ๋ฏธ๋””์–ด ๊ฒŒ์‹œ๋ฌผ์ด GTA6 ์ฝ”์ธ์ด๋‚˜ ์ด๋ฏธ ์‹œ์ž‘๋œ ์—์–ด๋“œ๋ž, ํฌ๋ฆฝํ† ๋กœ GTA6๋ฅผ ๋ฏธ๋ฆฌ ์‚ด ์ˆ˜ ์žˆ๋Š” ์‚ฌ์ „ ํŒ๋งค๋ฅผ ์•Œ๋ฆฐ๋‹ค. ๋งํฌ๋ฅผ ๋ˆ„๋ฅด๋ฉด ๋ฐ”์ด์Šค ์‹œํ‹ฐ ๊ทธ๋ฆผ์œผ๋กœ ๊พธ๋ฏผ ํŽ˜์ด์ง€๊ฐ€ ๋‚˜์˜ค๊ณ  ์ž๊ฒฉ ํ™•์ธ์ด๋‚˜ ์‚ฌ์ „ ์˜ˆ์•ฝ ๋ฒ„ํŠผ์„ ๋ˆ„๋ฅด๋ฉด ์ง€๊ฐ‘ ์—ฐ๊ฒฐ ์ฐฝ์ด ๋œฌ๋‹ค. ํƒˆ์ทจ๋Š” ์ด ์ง€๊ฐ‘ ์—ฐ๊ฒฐ ๋‹จ๊ณ„์—์„œ ์ผ์–ด๋‚œ๋‹ค. ๋ธ”๋ก์—์ด๋“œ์˜ ์†Œ์…œ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๊ธฐ์ค€์œผ๋กœ ์ตœ๊ทผ 4๊ฐœ์›” ๋™์•ˆ 116๊ฐœ ๊ณ„์ •์ด GTA ํ…Œ๋งˆ ๋ฏธ๋ผ ๊ฒŒ์‹œ๋ฌผ 451๊ฑด์„ ์˜ฌ๋ ธ๊ณ  ์ด ๊ฒŒ์‹œ๋ฌผ๋“ค์ด ์ˆ˜์‹ญ ๊ฐœ์˜ ์•…์„ฑ ๋””์•ฑ์œผ๋กœ ์—ฐ๊ฒฐ๋๋‹ค. ๋ฌธ๊ตฌ๋Š” ๊ฑฐ์˜ ๋ฐ”๋€Œ์ง€ ์•Š๋Š”๋‹ค. ํ•œ ์ค„์งœ๋ฆฌ ํ™๋ณด ๋ฌธ์žฅ๊ณผ ์—์–ด๋“œ๋ž์ด ์‹œ์ž‘๋์œผ๋‹ˆ ์ž๊ฒฉ์„ ํ™•์ธํ•˜๋ผ๋Š” ๋ง, ์งง์€ ๋งํฌ, ์บ์‹œํƒœ๊ทธ๊ฐ€ ๋ถ™๋Š”๋‹ค. ํŒ€์€ ์ ˆ๋Œ€ ๊ฐœ์ธ ๋งํฌ๋ฅผ ๋ณด๋‚ด์ง€ ์•Š๋Š”๋‹ค๋Š” ์•ˆ์ „ ์ˆ˜์น™์ฒ˜๋Ÿผ ๋“ค๋ฆฌ๋Š” ๋ฌธ์žฅ๋„ ๋ฐ˜๋ณตํ•ด์„œ ๋‚˜์˜ค๋Š”๋ฐ ์ด ์‚ฌ๊ธฐ์—์„œ๋Š” ๊ทธ ๋ฌธ์žฅ๊นŒ์ง€ ์†์ž„์ˆ˜์˜ ์ผ๋ถ€๋‹ค. ๋ฏธ๋ผ๋Š” 3๊ฐ€์ง€ ์ฒซ์งธ๋Š” ์‚ฌ์ „ ์˜ˆ์•ฝ ํŽ˜์ด์ง€์ธ๋ฐ 8์›” ์ค‘์ˆœ์— ๋‚˜์˜จ ์•…์„ฑ ๋””์•ฑ ๋ฌถ์Œ์€ ๋ชจ๋‘ Grand Theft Auto VI ยท Pre-order ยท November 19, 2026์ด๋ผ๋Š” ๊ฐ™์€ ์ œ๋ชฉ์„ ๋‹ฌ๊ณ  ์žˆ์—ˆ๋‹ค. ๋‘˜์งธ๋Š” ์–ผํ‹ฐ๋ฐ‹ ์—๋””์…˜ ํŒ๋งค ํŽ˜์ด์ง€์ธ๋ฐ ๊ฒฐ์ œ ๋‹จ๊ณ„๊ฐ€ ์นด๋“œ ๊ฒฐ์ œ๊ฐ€ ์•„๋‹ˆ๋ผ ์ง€๊ฐ‘ ์—ฐ๊ฒฐ์ด๋‹ค. ๋ธ”๋ก์—์ด๋“œ๊ฐ€ ์ฐพ์€ ์‚ฌ๋ก€ ํ•˜๋‚˜๋Š” ํ”Œ๋ ˆ์ด์Šคํ…Œ์ด์…˜ ์Šคํ† ์–ด๋ฅผ ๊ฑฐ์˜ ๊ทธ๋Œ€๋กœ ๋ฒ ๋‚€ ํŽ˜์ด์ง€๋กœ GTA6 ์–ผํ‹ฐ๋ฐ‹ ์—๋””์…˜์„ 89.99๋‹ฌ๋Ÿฌ์— ํŒ”๊ณ  ์žˆ์—ˆ๋‹ค. ์‚ฌ์ „ ์˜ˆ์•ฝ ๋ฒ„ํŠผ์„ ๋ˆ„๋ฅด๋ฉด ํŠธ๋ก ๋งํฌ์™€ ํŠธ๋Ÿฌ์ŠคํŠธ ์›”๋ ›, ๋ ˆ์ €๋ฅผ ๋น„๋กฏํ•œ 9๊ฐœ ์ง€๊ฐ‘ยท๊ฑฐ๋ž˜์†Œ ๋กœ๊ณ ๊ฐ€ ๋œจ๋Š” ์ง€๊ฐ‘ ์—ฐ๊ฒฐ ์ฐฝ์ด ๋‚˜์˜ค์ง€๋งŒ ์–ด๋А ๊ฒƒ์„ ๋ˆŒ๋Ÿฌ๋„ ๊ฐ™์€ ๋“œ๋ ˆ์ด๋„ˆ๋กœ ์ด์–ด์กŒ๋‹ค. ์…‹์งธ๋Š” ์—์–ด๋“œ๋ž๊ณผ ์–ผ๋ฆฌ ์•ก์„ธ์Šค ํŽ˜์ด์ง€๋กœ ์ดˆ๊ธฐ ์ฐธ์—ฌ์ž์—๊ฒŒ GTA6 ์ฝ”์ธ์„ ์ค€๋‹ค๋ฉด์„œ ์ž๊ฒฉ ํ™•์ธ ๋ฒ„ํŠผ๊ณผ ํด๋ ˆ์ž„ ๋ฒ„ํŠผ์„ ๋‹ฌ์•„ ๋‘”๋‹ค. ํŽ˜์ด์ง€ ๋’ค์˜ ํƒˆ์ทจ ๋„๊ตฌ๋Š” ๊ณ„์† ๋ฐ”๋€Œ์—ˆ๋Š”๋ฐ ๊ฐ™์€ ์–ผํ‹ฐ๋ฐ‹ ์—๋””์…˜ ๋ฏธ๋ผ๊ฐ€ 10์›” ๋ง์—๋Š” ํŠธ๋ก ๋งํฌ ์—ฐ๊ฒฐ ํ๋ฆ„์„ ์“ฐ๋Š” ๋‹ค๋ฅธ ๋“œ๋ ˆ์ด๋„ˆ ํ‚คํŠธ๋กœ ์˜ฎ๊ฒจ ๊ฐ”๊ณ  ๋‹ค๋ฅธ GTA ํŽ˜์ด์ง€๋“ค์€ ์†”๋ผ๋‚˜ ์—ฐ๊ฒฐ ํŒ์—… ํ‚คํŠธ๋ฅผ ์ผ๋‹ค. ๋ฏธ๋ผ๋Š” ์žฌ์‚ฌ์šฉํ•˜๊ณ  ๋’ค์— ๋ถ™๋Š” ๋„๊ตฌ๋งŒ ๊ฐˆ์•„ ๋ผ์šฐ๋Š” ๋ฐฉ์‹์ด๋‹ค. ํ† ํฐ์€ ๋Œ€๋ถ€๋ถ„ ๋ฏธ๋ผ ์—ญํ• ์ด๋‹ค 7์›” 4์ผ๋ถ€ํ„ฐ 11์›” 4์ผ๊นŒ์ง€ ์†”๋ผ๋‚˜์—์„œ GTA ์ด๋ฆ„์„ ๋‹จ ํ† ํฐ์ด 2๋งŒ ๊ฐœ ๋„˜๊ฒŒ ๋‚˜์™”๊ณ  ๋ธ”๋ก์—์ด๋“œ๋Š” ๊ทธ์ค‘ 8800๊ฐœ ์ด์ƒ์„ ์•…์„ฑ์ด๋‚˜ ๊ฒฝ๊ณ , ์ŠคํŒธ์œผ๋กœ ๋ถ„๋ฅ˜ํ–ˆ๋‹ค. ๋ถ„๋ฅ˜๋œ ํ† ํฐ์˜ 46%๋Š” ํŽŒํ”„ํŽ€ ํ˜•ํƒœ์˜ ๋Ÿฐ์น˜ํŒจ๋“œ์—์„œ ๋‚˜์™”๋‹ค. EVM ์ฒด์ธ์—์„œ๋Š” ๋” ์—„๊ฒฉํ•œ ์ด๋ฆ„ ๊ธฐ์ค€์œผ๋กœ ๋ฒ ์ด์Šค 593๊ฐœ, ์ด๋”๋ฆฌ์›€ 111๊ฐœ๊ฐ€ ์žกํ˜”๋‹ค. ์ด ํ† ํฐ๋“ค ๋Œ€๋ถ€๋ถ„์€ ํƒˆ์ทจ ์ˆ˜๋‹จ์ด๋ผ๊ธฐ๋ณด๋‹ค ๋ฏธ๋ผ๋‹ค. ๊ฒŒ์‹œ๋ฌผ์—์„œ $GTA6๋‚˜ $GTAVI ์˜†์— ๋ถ™์€ ์†”๋ผ๋‚˜ ์ฃผ์†Œ๋Š” ๋Œ€๊ฐœ ๊ด€๊ณ„์—†๋Š” ์ธ๊ธฐ ๋ฐˆ์ฝ”์ธ ์ฃผ์†Œ์˜€๊ณ  ๊ทธ๋Ÿด๋“ฏํ•ด ๋ณด์ด๊ฒŒ ๋งŒ๋“œ๋Š” ์šฉ๋„์˜€๋‹ค. ๋‹ค๋งŒ ํ† ํฐ ์ž์ฒด๊ฐ€ ํ•จ์ •์ธ ๊ฒฝ์šฐ๋„ ์žˆ๋‹ค. 11์›” 3์ผ ํ•œ ๊ตฌ๋งค์ž๋Š” ์œ ๋‹ˆ์Šค์™‘์—์„œ GTA6DOG(RICO)๋ฅผ ์ƒ€๋‹ค๊ฐ€ ์•ฝ 1๋ถ„ ๋งŒ์— ํ† ํฐ ๋ฐฐํฌ์ž๊ฐ€ ์ปจํŠธ๋ž™ํŠธ์˜ transferFrom์„ ํ˜ธ์ถœํ•ด ๊ตฌ๋งค์ž์˜ RICO 27๋งŒ1031๊ฐœ ์ „๋ถ€๋ฅผ ์†Œ๊ฐ ์ฃผ์†Œ๋กœ ๋ณด๋‚ด๋Š” ์ผ์„ ๊ฒช์—ˆ๋‹ค. ๊ตฌ๋งค์ž ์ง€๊ฐ‘์˜ ์Šน์ธ์€ ์—†์—ˆ๋‹ค. ์ปจํŠธ๋ž™ํŠธ๋Š” ๋งŒ๋“ค์–ด์ง„ ์ง€ ์•ฝ 11์‹œ๊ฐ„ ๋œ ์ƒํƒœ์˜€๊ณ  ์•ž์„œ ์†Œ์•ก ๋งค๋„ ํ…Œ์ŠคํŠธ๊ฐ€ ํ†ต๊ณผ๋ผ์„œ ์ •์ƒ์ ์œผ๋กœ ๊ฑฐ๋ž˜๋˜๋Š” ํ† ํฐ์ฒ˜๋Ÿผ ๋ณด์˜€๋‹ค. ๋ธ”๋ก์—์ด๋“œ ํ† ํฐ ์Šค์บ๋„ˆ๋Š” 11์›” 3์ผ ์ฒซ ์Šค์บ” ํ›„ 1์‹œ๊ฐ„ ์•ˆ์— ์ด ํ† ํฐ์„ ์œ„ํ—˜์œผ๋กœ ํ‘œ์‹œํ–ˆ๊ณ  ๊ณต๊ธ‰๋Ÿ‰ ๋Œ€๋ถ€๋ถ„์„ ํ•œ ํ‚ค ๋ณด์œ ์ž๊ฐ€ ์—ฌ๋Ÿฌ ์ฐจ๋ช… ์ง€๊ฐ‘์œผ๋กœ ์ฅ๊ณ  ์žˆ๋‹ค๊ณ  ํŒ๋‹จํ–ˆ๋‹ค. ๋ธ”๋ก์—์ด๋“œ๊ฐ€ ๋ง‰๋Š” ๋ฐฉ์‹ ๋ธ”๋ก์—์ด๋“œ๋Š” ์‚ฌ์ดํŠธ ์ฝ”๋“œ๋ฅผ ์ด๋ฏธ ์•Œ๋ ค์ง„ ๋“œ๋ ˆ์ด๋„ˆ ํ‚คํŠธ์™€ ๋Œ€์กฐํ•ด์„œ ์ƒˆ ๋„๋ฉ”์ธ์ด๋ผ๋„ ๊ฐ™์€ ํ‚คํŠธ๋ฅผ ์“ฐ๋ฉด ์ฒซ ์Šค์บ”์—์„œ ๊ฑธ๋Ÿฌ ๋‚ธ๋‹ค. ์„œ๋ช… ์ „์—๋Š” ๊ฑฐ๋ž˜๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•ด์„œ ํด๋ ˆ์ž„์ด๋ผ๊ณ  ์ ํžŒ ๋ฒ„ํŠผ์ด ์‹ค์ œ๋กœ๋Š” ์ „์†ก์ด๋‚˜ ๋ฌด์ œํ•œ ์Šน์ธ์ธ ๊ฒฝ์šฐ๋ฅผ ๋“œ๋Ÿฌ๋‚ธ๋‹ค. ์ด ํŒ์ •์€ ๋ธ”๋ก์—์ด๋“œ๋ฅผ ์—ฐ๋™ํ•œ ์ง€๊ฐ‘๊ณผ ๊ฑฐ๋ž˜์†Œ, ๋””์•ฑ์—์„œ ์—ฐ๊ฒฐ์ด๋‚˜ ์„œ๋ช… ์ˆœ๊ฐ„์˜ ๊ฒฝ๊ณ ๋‚˜ ์ฐจ๋‹จ์œผ๋กœ ๋‚˜ํƒ€๋‚œ๋‹ค. ๋ธ”๋ก์—์ด๋“œ๊ฐ€ ์ •๋ฆฌํ•œ ์ฒซ ๊ธฐ์ค€์€ ๊ฒŒ์ž„์„ ๊ณต์‹ ์Šคํ† ์–ด์—์„œ๋งŒ ์‚ฌ๋Š” ๊ฒƒ์ด๋‹ค. ๋ฝ์Šคํƒ€์™€ ํ”Œ๋ ˆ์ด์Šคํ…Œ์ด์…˜, ์—‘์Šค๋ฐ•์Šค, ์ŠคํŒ€์€ GTA6 ๊ฒฐ์ œ์— ํฌ๋ฆฝํ† ๋ฅผ ๋ฐ›์ง€ ์•Š๊ณ  ์ง€๊ฐ‘ ์—ฐ๊ฒฐ๋„ ์š”๊ตฌํ•˜์ง€ ์•Š๋Š”๋‹ค. ๊ณต์‹ GTA ํ† ํฐ์ด๋‚˜ ์—์–ด๋“œ๋ž๋„ ์—†๋‹ค. ์ž๊ฒฉ์„ ํ™•์ธํ•˜๋ผ๋Š” ๋ฌธ๊ตฌ๋Š” ๊ทธ ์ž์ฒด๋กœ ๊ฒฝ๊ณ  ์‹ ํ˜ธ๋‹ค. ๊ฒŒ์ž„์„ ํŒ”๊ธฐ ์œ„ํ•ด ์ง€๊ฐ‘์„ ๋ณผ ํ•„์š”๊ฐ€ ์žˆ๋Š” ํŒ๋งค์ฒ˜๋Š” ์—†๋‹ค. ์ง€๊ฐ‘์ด ์•…์„ฑ ์‚ฌ์ดํŠธ๋ผ๊ณ  ๊ฒฝ๊ณ ํ•˜๊ฑฐ๋‚˜ ํด๋ ˆ์ž„ ๋ฒ„ํŠผ์ด ๋‚ด ํ† ํฐ์„ ์˜ฎ๊ธฐ๋Š” ๊ฑฐ๋ž˜๋กœ ๋‚˜์˜ค๋ฉด ๊ฑฐ๊ธฐ์„œ ๋ฉˆ์ถ˜๋‹ค. ์ต์ˆ™ํ•œ ์ง€๊ฐ‘ ๋กœ๊ณ ๊ฐ€ ๋–  ์žˆ๋‹ค๊ณ  ์•ˆ์‹ฌํ•  ์ด์œ ๋„ ์—†๋‹ค. ์ด๋ฒˆ ๊ฐ€์งœ ์—ฐ๊ฒฐ ์ฐฝ์€ ํŠธ๋Ÿฌ์ŠคํŠธ ์›”๋ ›๊ณผ ๋ ˆ์ €, OKX ๋กœ๊ณ ๋ฅผ ๊ทธ๋Œ€๋กœ ์ผ๋‹ค.
2
1
91
AI๋กœ ๊ธ€ ์“ฐ๋Š” ์‚ฌ๋žŒ ์ฃผ๋ชฉ ํ•˜์…ˆ GPT ์ดˆ์•ˆ ๊ทธ๋Œ€๋กœ ๋ณต๋ถ™ํ•˜์ง€ ๋ง๊ณ  AI ํ‹ฐ ์‹น ์ง€์›Œ์ฃผ๋Š” ํ•œ๊ตญ์–ด ๊ธ€์“ฐ๊ธฐ ์Šคํ‚ฌ 4๊ฐœ ๋“ฑ๋กํ•ด์„œ ์“ฐ์…ˆ 1. humanizer-ko (ํ•œ๊ตญ์–ด AI ํ‹ฐ ์ œ๊ฑฐ) ๋‹ค์–‘ํ•œ ํ•ต์‹ฌ์ ์ธ ๊ฐ™์€ AI ๋ง๋ฒ„๋ฆ‡์ด๋ž‘ ๋ฒˆ์—ญํˆฌ ์ด์ค‘ ํ”ผ๋™๊นŒ์ง€ 37๊ฐ€์ง€ ํŒจํ„ด์œผ๋กœ ์žก์•„์คŒ ๊ณ ์น  ๊ฒŒ ์—†์œผ๋ฉด ์•ˆ ๊ณ ์นจ ๐Ÿ”— github.com/IsaacEryn/humanizโ€ฆ 2. korean-prose-skill (๋ฒˆ์—ญํˆฌ ๋ฌธ์žฅ ๋‹ค๋“ฌ๊ธฐ) ์˜์–ด ์ง์—ญํ•œ ๊ธฐ์ˆ  ๋ฌธ์„œ PR ๋ณธ๋ฌธ UI ๋ฌธ๊ตฌ๋ฅผ ํ•œ๊ตญ ์‚ฌ๋žŒ์ด ์ฝ๊ธฐ ์ž์—ฐ์Šค๋Ÿฌ์šด ๋ฌธ์žฅ์œผ๋กœ ๋ฐ”๊ฟ”์คŒ ๐Ÿ”— github.com/JellyBrick/koreanโ€ฆ 3. korean-skills (๋งž์ถค๋ฒ• ๋„์–ด์“ฐ๊ธฐ ๊ฒ€์‚ฌ) ํ•œ๊ตญ์–ด ๋งž์ถค๋ฒ• ๋„์–ด์“ฐ๊ธฐ ๋ฌธ๋ฒ• ๊ฒ€์‚ฌ์— ๋ฌธ์„œ ์Šคํƒ€์ผ ํ†ต์ผ๊นŒ์ง€ ์Šคํ‚ฌ 3๊ฐœ๋กœ ํ•œ ๋ฒˆ์— ํ•ด์คŒ ๐Ÿ”— github.com/DaleSeo/korean-skโ€ฆ 4. academic-research-skills (๋…ผ๋ฌธ ์ž‘์„ฑ ๋„์šฐ๋ฏธ) ์—ฐ๊ตฌ ์„ค๊ณ„๋ถ€ํ„ฐ ๋…ผ๋ฌธ ์ž‘์„ฑ ๊ฒ€ํ†  ์ถœํŒ ์ค€๋น„๊นŒ์ง€ ๋‹จ๊ณ„๋งˆ๋‹ค ํ™•์ธ๋ฐ›์œผ๋ฉด์„œ ๊ฐ™์ด ์ง„ํ–‰ํ•ด์คŒ ๐Ÿ”— github.com/Imbad0202/academiโ€ฆ ๋งํฌ ๋‹ค ๋‚จ๊ฒจ๋’€์œผ๋‹ˆ ์ €์žฅํ•ด๋‘๊ณ  ์จ๋จน์œผ์…ˆ
60
2,245
96
5,547
518,112
๐Ÿ“š Rialo Research Series 66: Agent State Integrity - Can a Recovered Agent Trust Its Own Memory? (Kor : ์—์ด์ „ํŠธ ์ƒํƒœ ๋ฌด๊ฒฐ์„ฑ - ๋ณต๊ตฌ๋œ Agent๋Š” ์ž์‹ ์˜ ๊ธฐ์–ต์„ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ์„๊นŒ?) 1/ Intro: Recovery Doesn't Guarantee Integrity (EN) In the previous chapter, we explored how a suspended autonomous agent could safely regain execution authority. But restoring permissions doesn't necessarily mean the agent's internal state is trustworthy. Imagine a lending agent suspended after receiving manipulated oracle prices. The data source is corrected, its permissions are reviewed, and execution resumes. However, the agent still remembers an incorrect liquidation threshold calculated before the shutdown. An agent can be operationally recovered while remaining logically compromised. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ์ค‘๋‹จ๋œ Autonomous Agent๊ฐ€ ์–ด๋–ป๊ฒŒ ์•ˆ์ „ํ•˜๊ฒŒ Execution Authority๋ฅผ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ๊ถŒํ•œ์„ ๋ณต๊ตฌํ–ˆ๋‹ค๊ณ  ํ•ด์„œ Agent์˜ Internal State๊นŒ์ง€ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์กฐ์ž‘๋œ Oracle ๊ฐ€๊ฒฉ์„ ๋ฐ›์•„ ์ค‘๋‹จ๋œ Lending Agent๋ฅผ ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๊ฐ€ ์ˆ˜์ •๋˜๊ณ  Permission Review๊นŒ์ง€ ์™„๋ฃŒ๋ผ ์‹คํ–‰์„ ์žฌ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Agent ๋‚ด๋ถ€์—๋Š” ์ค‘๋‹จ ์ด์ „์— ๊ณ„์‚ฐํ•œ ์ž˜๋ชป๋œ Liquidation Threshold๊ฐ€ ์—ฌ์ „ํžˆ ๋‚จ์•„ ์žˆ์Šต๋‹ˆ๋‹ค. Agent๋Š” ์ •์ƒ์ ์œผ๋กœ ์žฌ๊ฐ€๋™๋๋”๋ผ๋„ ๋‚ด๋ถ€ ํŒ๋‹จ ์ƒํƒœ๋Š” ์—ฌ์ „ํžˆ ์˜ค์—ผ๋ผ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 2/ What Makes Up an Agent's State? (EN) An autonomous agent's state may contain more than wallet balances or transaction history. It can include cached market data, risk calculations, execution queues, temporary permissions, and previous decisions. Some information may be stored on-chain, while other components remain in off-chain databases or agent memory. Verifying only blockchain transactions therefore cannot establish the integrity of the entire agent. A reliable recovery process needs to identify which state components influenced the failed decision. (KR) Autonomous Agent์˜ State๋Š” Wallet Balance๋‚˜ Transaction History๋งŒ์œผ๋กœ ๊ตฌ์„ฑ๋˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. Cached Market Data, Risk Calculation, Execution Queue, Temporary Permission, Previous Decision ๋“ฑ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ ์ •๋ณด๋Š” On-chain์— ๊ธฐ๋ก๋˜์ง€๋งŒ ๋‹ค๋ฅธ ์ •๋ณด๋Š” Off-chain Database๋‚˜ Agent Memory์— ๋‚จ์•„ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Blockchain Transaction๋งŒ ๊ฒ€์ฆํ•œ๋‹ค๊ณ  Agent ์ „์ฒด์˜ ๋ฌด๊ฒฐ์„ฑ์ด ๋ณด์žฅ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์‹คํŒจํ•œ ํŒ๋‹จ์— ์–ด๋–ค State Component๊ฐ€ ์˜ํ–ฅ์„ ๋ฏธ์ณค๋Š”์ง€ ํ™•์ธํ•˜๋Š” ๊ณผ์ •์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 3/ The Danger of Contaminated Memory (EN) Suppose an agent stores a summary stating that a particular asset has become dangerously undercollateralized. That summary was generated from manipulated prices. Even after the oracle is corrected, the agent might continue using the stored conclusion instead of recalculating the risk. This becomes especially problematic when agents rely on persistent memory to make decisions across multiple sessions. The original data may be fixed, but the incorrect interpretation can survive. (KR) Agent๊ฐ€ ํŠน์ • ์ž์‚ฐ์˜ ๋‹ด๋ณด ๋น„์œจ์ด ์œ„ํ—˜ ์ˆ˜์ค€์œผ๋กœ ๋–จ์–ด์กŒ๋‹ค๋Š” Summary๋ฅผ ์ €์žฅํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ํ•ด๋‹น Summary๋Š” ์กฐ์ž‘๋œ ๊ฐ€๊ฒฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋์Šต๋‹ˆ๋‹ค. Oracle์ด ์ˆ˜์ •๋œ ์ดํ›„์—๋„ Agent๊ฐ€ Risk๋ฅผ ๋‹ค์‹œ ๊ณ„์‚ฐํ•˜์ง€ ์•Š๊ณ  ๊ธฐ์กด ๊ฒฐ๋ก ์„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ ์—ฌ๋Ÿฌ Session์— ๊ฑธ์ณ Persistent Memory๋ฅผ ์‚ฌ์šฉํ•˜๋Š” Agent๋ผ๋ฉด ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ๋ณต์žกํ•ด์ง‘๋‹ˆ๋‹ค. ์›๋ณธ ๋ฐ์ดํ„ฐ๋Š” ์ˆ˜์ •๋์ง€๋งŒ ์ž˜๋ชป๋œ ํ•ด์„์€ Memory์— ๊ณ„์† ๋‚จ์•„ ์žˆ์„ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. 4/ State Reconstruction and Verification (EN) One possible approach is reconstructing critical state from verified inputs. The application could identify the last trusted checkpoint, retrieve authenticated data, and recompute affected risk assessments. Stored decisions would then be compared with newly calculated results. Any mismatch could trigger further investigation before execution resumes. However, reconstructing state requires access to reliable historical inputs and deterministic processing where applicable. A checkpoint is only useful if its own integrity can be established. (KR) ํ•œ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์€ ๊ฒ€์ฆ๋œ Input์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ค‘์š”ํ•œ State๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. Application์€ ๋งˆ์ง€๋ง‰์œผ๋กœ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” Checkpoint๋ฅผ ํ™•์ธํ•˜๊ณ , ์ธ์ฆ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ถˆ๋Ÿฌ์™€ ์˜ํ–ฅ์„ ๋ฐ›์€ Risk Assessment๋ฅผ ๋‹ค์‹œ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ์ €์žฅ๋œ Decision๊ณผ ์ƒˆ๋กญ๊ฒŒ ๊ณ„์‚ฐํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๋ถˆ์ผ์น˜๊ฐ€ ๋ฐœ๊ฒฌ๋˜๋ฉด ์‹คํ–‰ ์žฌ๊ฐœ ์ „์— ์ถ”๊ฐ€ ๊ฒ€์ฆ์„ ์ง„ํ–‰ํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๋‹ค๋งŒ State Reconstruction์—๋Š” ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” Historical Input๊ณผ ํ•„์š”ํ•œ ๊ฒฝ์šฐ ์žฌํ˜„ ๊ฐ€๋Šฅํ•œ ๊ณ„์‚ฐ ๊ณผ์ •์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. Checkpoint ์—ญ์‹œ ๋ฌด๊ฒฐ์„ฑ์ด ํ™•์ธ๋ผ์•ผ ๋ณต๊ตฌ ๊ธฐ์ค€์œผ๋กœ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 5/ When State Cannot Be Fully Reconstructed (EN) Not every agent decision can be perfectly reproduced. An AI model may generate different outputs when its model version, context, or external dependencies change. Some off-chain information may also be unavailable after an incident. In these situations, pretending that a complete reconstruction is possible would create false confidence. Applications may need to invalidate uncertain state, restrict sensitive actions, and require additional authorization before restoring execution. Unverifiable state should not automatically become trusted state. (KR) ๋ชจ๋“  Agent์˜ ํŒ๋‹จ์„ ์™„๋ฒฝํ•˜๊ฒŒ ์žฌํ˜„ํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. AI Model Version์ด๋‚˜ Context, External Dependency๊ฐ€ ๋ณ€๊ฒฝ๋˜๋ฉด ๊ฐ™์€ ์ž…๋ ฅ์—์„œ๋„ ๋‹ค๋ฅธ ๊ฒฐ๊ณผ๊ฐ€ ์ƒ์„ฑ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ผ๋ถ€ Off-chain Information์€ ์‚ฌ๊ณ  ์ดํ›„ ํ™•๋ณดํ•˜์ง€ ๋ชปํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฐ ์ƒํ™ฉ์—์„œ ์™„์ „ํ•œ Reconstruction์ด ๊ฐ€๋Šฅํ•˜๋‹ค๊ณ  ๊ฐ€์ •ํ•˜๋ฉด ์˜คํžˆ๋ ค ์ž˜๋ชป๋œ ์‹ ๋ขฐ๊ฐ€ ํ˜•์„ฑ๋ฉ๋‹ˆ๋‹ค. Application์€ ๊ฒ€์ฆํ•  ์ˆ˜ ์—†๋Š” State๋ฅผ ๋ฌดํšจํ™”ํ•˜๊ฑฐ๋‚˜ ๊ณ ์œ„ํ—˜ ํ–‰๋™์„ ์ œํ•œํ•˜๊ณ , ์‹คํ–‰ ๊ถŒํ•œ ๋ณต๊ตฌ์— ์ถ”๊ฐ€ ์Šน์ธ์„ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒ€์ฆํ•  ์ˆ˜ ์—†๋Š” ์ƒํƒœ๋ฅผ ๋‹จ์ˆœํžˆ ์ •์ƒ ์ƒํƒœ๋กœ ๊ฐ„์ฃผํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค. 6/ Rialo and State Integrity (EN) Rialo's execution-oriented architecture offers an interesting environment for exploring state integrity checks. Reactive Transactions could help applications trigger verification workflows when predefined inconsistencies are detected. Execution records and application-managed checkpoints could support investigations into how an agent's state changed. However, Rialo's reactive capabilities do not automatically verify arbitrary off-chain memory or guarantee deterministic AI outputs. Developers must define which state elements require verification, how trusted references are established, and what happens when validation fails. (KR) Rialo์˜ Execution ์ค‘์‹ฌ ์•„ํ‚คํ…์ฒ˜๋Š” State Integrity ๊ฒ€์ฆ์„ ํƒ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ํ™˜๊ฒฝ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Reactive Transactions๋ฅผ ํ™œ์šฉํ•˜๋ฉด ์‚ฌ์ „์— ์ •์˜๋œ ๋ถˆ์ผ์น˜๊ฐ€ ๊ฐ์ง€๋์„ ๋•Œ Verification Workflow๋ฅผ ์‹œ์ž‘ํ•˜๋„๋ก ์„ค๊ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Execution Record์™€ Application์ด ๊ด€๋ฆฌํ•˜๋Š” Checkpoint๋Š” Agent State๊ฐ€ ์–ด๋–ป๊ฒŒ ๋ณ€๊ฒฝ๋๋Š”์ง€ ์กฐ์‚ฌํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ Rialo์˜ Reactive ๊ธฐ๋Šฅ์ด ๋ชจ๋“  Off-chain Memory๋ฅผ ์ž๋™์œผ๋กœ ๊ฒ€์ฆํ•˜๊ฑฐ๋‚˜ AI Output์˜ ๊ฒฐ์ •์„ฑ์„ ๋ณด์žฅํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์–ด๋–ค State๋ฅผ ๊ฒ€์ฆํ•˜๊ณ , ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ์ค€์„ ์–ด๋–ป๊ฒŒ ์„ค์ •ํ•˜๋ฉฐ, ๊ฒ€์ฆ ์‹คํŒจ ์‹œ ์–ด๋–ค ์กฐ์น˜๋ฅผ ์ทจํ• ์ง€๋Š” ๊ฐœ๋ฐœ์ž๊ฐ€ ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a demonstration where a lending agent is suspended after using manipulated oracle data. The oracle is corrected, and the agent requests permission to resume. Before approval, the application compares the agent's stored risk assessment against a newly reconstructed state. The verification process discovers that the agent's memory still contains an incorrect liquidation threshold. Execution remains restricted until the contaminated state is invalidated and the calculation is repeated. This would demonstrate why restoring execution authority and restoring trustworthy decision-making are separate requirements. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์กฐ์ž‘๋œ Oracle ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ ๋’ค ์ค‘๋‹จ๋œ Lending Agent์˜ ๋ณต๊ตฌ Demo์ž…๋‹ˆ๋‹ค. Oracle์ด ์ •์ƒํ™”๋˜๊ณ  Agent๊ฐ€ ์‹คํ–‰ ์žฌ๊ฐœ๋ฅผ ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์Šน์ธ ์ „์— Application์ด ์ €์žฅ๋œ Risk Assessment์™€ ์ƒˆ๋กญ๊ฒŒ ์žฌ๊ตฌ์„ฑํ•œ State๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ฆ ๊ณผ์ •์—์„œ Agent Memory์— ์ž˜๋ชป๋œ Liquidation Threshold๊ฐ€ ๋‚จ์•„ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ๋ฐœ๊ฒฌ๋ฉ๋‹ˆ๋‹ค. ์˜ค์—ผ๋œ State๋ฅผ ๋ฌดํšจํ™”ํ•˜๊ณ  ๊ณ„์‚ฐ์„ ๋‹ค์‹œ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์ „๊นŒ์ง€ Execution์€ ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ์‹คํ–‰ ๊ถŒํ•œ ๋ณต๊ตฌ์™€ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ํŒ๋‹จ ๋Šฅ๋ ฅ์˜ ๋ณต๊ตฌ๊ฐ€ ์„œ๋กœ ๋‹ค๋ฅธ ์กฐ๊ฑด์ด๋ผ๋Š” ์ ์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. (EN) Autonomous finance increasingly depends on agents that remember previous decisions and operate continuously across multiple environments. Persistent memory improves efficiency, but it can also preserve mistakes long after their original causes disappear. A trustworthy recovery process must therefore examine not only what an agent is allowed to do, but also what information it still believes. An autonomous agent should not regain financial authority simply because it can run again. Its decision state must be trustworthy enough to justify that authority. (KR) Autonomous Finance์—์„œ๋Š” ์ด์ „ ํŒ๋‹จ์„ ๊ธฐ์–ตํ•˜๊ณ  ์—ฌ๋Ÿฌ ํ™˜๊ฒฝ์—์„œ ์ง€์†์ ์œผ๋กœ ์ž‘๋™ํ•˜๋Š” Agent์˜ ์ค‘์š”์„ฑ์ด ์ปค์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Persistent Memory๋Š” ํšจ์œจ์„ฑ์„ ๋†’์ด์ง€๋งŒ, ์ตœ์ดˆ ์˜ค๋ฅ˜๊ฐ€ ์‚ฌ๋ผ์ง„ ์ดํ›„์—๋„ ์ž˜๋ชป๋œ ํŒ๋‹จ์„ ๋ณด์กดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” Recovery Process๋ผ๋ฉด Agent๊ฐ€ ๋ฌด์—‡์„ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋งŒ ํ™•์ธํ•ด์„œ๋Š” ๋ถ€์กฑํ•ฉ๋‹ˆ๋‹ค. Agent๊ฐ€ ์–ด๋–ค ์ •๋ณด๋ฅผ ์‚ฌ์‹ค๋กœ ์ธ์‹ํ•˜๊ณ  ์žˆ๋Š”์ง€๊นŒ์ง€ ๊ฒ€์ฆํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ Autonomous Finance์—์„œ ์ค‘์š”ํ•œ ๊ฒƒ์€ Agent๋ฅผ ๋‹ค์‹œ ์ž‘๋™์‹œํ‚ค๋Š” ๊ฒƒ๋ฟ ์•„๋‹ˆ๋ผ, ๊ทธ Agent๊ฐ€ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ƒํƒœ์—์„œ ๋‹ค์‹œ ํŒ๋‹จํ•˜๋„๋ก ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ์•„๋‹๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
๐Ÿ“š Rialo Research Series 65: Agent Recovery Authorization - When Is It Safe to Restart an Autonomous Agent? (Kor : ์—์ด์ „ํŠธ ๋ณต๊ตฌ ์Šน์ธ - ์ค‘๋‹จ๋œ ์ž์œจ Agent๋ฅผ ์–ธ์ œ ๋‹ค์‹œ ์‹คํ–‰ํ•ด๋„ ์•ˆ์ „ํ• ๊นŒ?) 1/ Intro: Stopping an Agent Is Only Half the Solution (EN) In the previous chapter, we explored how emergency shutdown authority could prevent an autonomous agent from causing further financial damage. But suspending execution creates another problem. Imagine a lending agent that was stopped after detecting inconsistent oracle prices. The oracle eventually returns to normal, but should the agent immediately resume its operations? What if its pending instructions, permissions, or internal state still reflect the corrupted data? Stopping an unsafe agent is one challenge. Proving that it is safe to restart is another. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” Emergency Shutdown Authority๋ฅผ ํ†ตํ•ด ์œ„ํ—˜ํ•œ Agent์˜ ์ถ”๊ฐ€์ ์ธ ๊ธˆ์œต ํ”ผํ•ด๋ฅผ ๋ง‰๋Š” ๋ฐฉ๋ฒ•์„ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์‹คํ–‰์„ ์ค‘๋‹จํ•˜๋ฉด ๋˜ ๋‹ค๋ฅธ ๋ฌธ์ œ๊ฐ€ ์ƒ๊น๋‹ˆ๋‹ค. Oracle ๊ฐ€๊ฒฉ ๋ถˆ์ผ์น˜๊ฐ€ ๊ฐ์ง€๋ผ Lending Agent๊ฐ€ ์ค‘๋‹จ๋๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ดํ›„ Oracle์ด ์ •์ƒํ™”๋๋‹ค๊ณ  ํ•ด์„œ Agent๋ฅผ ์ฆ‰์‹œ ์žฌ๊ฐ€๋™ํ•ด๋„ ๊ดœ์ฐฎ์„๊นŒ์š”? ๊ธฐ์กด Pending Instruction์ด๋‚˜ Permission, Internal State์— ์—ฌ์ „ํžˆ ์ž˜๋ชป๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ฐ˜์˜๋ผ ์žˆ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„ํ—˜ํ•œ Agent๋ฅผ ์ค‘๋‹จํ•˜๋Š” ๊ฒƒ๊ณผ ์•ˆ์ „ํ•˜๊ฒŒ ๋‹ค์‹œ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Œ์„ ์ฆ๋ช…ํ•˜๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. 2/ Why Normal Data Doesn't Mean Normal Operation (EN) Suppose an agent calculates liquidation risk using a corrupted collateral price. The oracle is corrected, but the agent has already stored an incorrect risk assessment. If execution resumes without refreshing that assessment, the agent could submit a liquidation based on outdated information. Restoring the original data source does not automatically repair every downstream decision. Recovery must verify both the external inputs and the internal state derived from them. (KR) Agent๊ฐ€ ์ž˜๋ชป๋œ ๋‹ด๋ณด ๊ฐ€๊ฒฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ Liquidation Risk๋ฅผ ๊ณ„์‚ฐํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. Oracle ๊ฐ€๊ฒฉ์€ ์ˆ˜์ •๋์ง€๋งŒ Agent ๋‚ด๋ถ€์—๋Š” ์ด๋ฏธ ์ž˜๋ชป๋œ Risk Assessment๊ฐ€ ์ €์žฅ๋ผ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒํƒœ์—์„œ ์‹คํ–‰์„ ์žฌ๊ฐœํ•˜๋ฉด ์˜ค๋ž˜๋œ ํŒ๋‹จ์„ ๊ธฐ๋ฐ˜์œผ๋กœ Liquidation Transaction์„ ์ œ์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ๊ฐ€ ์ •์ƒํ™”๋๋‹ค๊ณ  ํ•ด์„œ ๊ทธ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋œ ๋ชจ๋“  ํŒ๋‹จ๊นŒ์ง€ ๋ณต๊ตฌ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. Recovery ๊ณผ์ •์—์„œ๋Š” External Input๊ณผ Internal State๋ฅผ ํ•จ๊ป˜ ๊ฒ€์ฆํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 3/ Recovery Preconditions (EN) An agent should not restart simply because a timeout has expired or an administrator presses a button. Applications could define explicit recovery conditions. For example, oracle prices must remain consistent across multiple observations, affected risk calculations must be refreshed, and unsafe pending instructions must be invalidated. Previously granted permissions should also be reviewed. Only after these conditions are satisfied should the system consider restoring execution authority. (KR) ์ผ์ • ์‹œ๊ฐ„์ด ์ง€๋‚ฌ๊ฑฐ๋‚˜ ๊ด€๋ฆฌ์ž๊ฐ€ ๋ฒ„ํŠผ์„ ๋ˆŒ๋ €๋‹ค๋Š” ์ด์œ ๋งŒ์œผ๋กœ Agent๋ฅผ ์žฌ๊ฐ€๋™ํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค. Application์€ ๋ช…ํ™•ํ•œ Recovery Condition์„ ์ •์˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์—ฌ๋Ÿฌ ์ฐจ๋ก€ ๊ด€์ธก์—์„œ Oracle ๊ฐ€๊ฒฉ์ด ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•ด์•ผ ํ•˜๊ณ , ์˜ํ–ฅ์„ ๋ฐ›์€ Risk Calculation์„ ๋‹ค์‹œ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์œ„ํ—˜ํ•œ Pending Instruction์€ ๋ฌดํšจํ™”ํ•˜๊ณ  ๊ธฐ์กด Permission๋„ ์žฌ๊ฒ€ํ† ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์กฐ๊ฑด์ด ์ถฉ์กฑ๋œ ์ดํ›„์— ์‹คํ–‰ ๊ถŒํ•œ ๋ณต๊ตฌ๋ฅผ ๊ฒ€ํ† ํ•˜๋Š” ๊ตฌ์กฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 4/ Who Approves the Restart? (EN) The authority to suspend an agent does not necessarily need to be identical to the authority to restart it. Emergency intervention prioritizes speed. Recovery authorization prioritizes verification. A security operator might be allowed to suspend execution immediately, while restarting high-risk operations requires additional approval. For example, a multi-signature authorization process could prevent one compromised administrator from restoring unsafe permissions. However, the approval structure must remain practical enough to avoid unnecessary service disruption. (KR) Agent๋ฅผ ์ค‘๋‹จํ•  ์ˆ˜ ์žˆ๋Š” ๊ถŒํ•œ๊ณผ ๋‹ค์‹œ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ถŒํ•œ์ด ๋ฐ˜๋“œ์‹œ ๊ฐ™์„ ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. Emergency Intervention์—์„œ๋Š” ๋Œ€์‘ ์†๋„๊ฐ€ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด Recovery Authorization์—์„œ๋Š” ๊ฒ€์ฆ ๊ณผ์ •์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์šด์˜์ž๊ฐ€ ์ฆ‰์‹œ ์‹คํ–‰์„ ์ค‘๋‹จํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋˜, ๊ณ ์œ„ํ—˜ ๊ธฐ๋Šฅ์˜ ์žฌ๊ฐ€๋™์—๋Š” ์ถ”๊ฐ€ ์Šน์ธ์„ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Multi-signature Authorization์„ ํ™œ์šฉํ•˜๋ฉด ํƒˆ์ทจ๋œ ๊ด€๋ฆฌ์ž ํ•œ ๋ช…์ด ์œ„ํ—˜ํ•œ ๊ถŒํ•œ์„ ์ž„์˜๋กœ ๋ณต๊ตฌํ•˜๋Š” ์ƒํ™ฉ์„ ๋ฐฉ์ง€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋‹ค๋งŒ ์Šน์ธ ์ ˆ์ฐจ๊ฐ€ ์ง€๋‚˜์น˜๊ฒŒ ๋ณต์žกํ•ด ์ •์ƒ์ ์ธ ์„œ๋น„์Šค ๋ณต๊ตฌ๊นŒ์ง€ ๋ฐฉํ•ดํ•˜์ง€ ์•Š๋„๋ก ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 5/ Gradual Recovery Instead of Full Restart (EN) Recovery does not have to be an all-or-nothing decision. An agent could first resume read-only operations and market monitoring without permission to execute financial transactions. After its calculations are verified, limited transaction capabilities could be restored under stricter limits. Full execution authority would return only after additional checks. This staged approach reduces the risk of immediately repeating the original failure. Recovery should restore capabilities according to verified safety, not merely elapsed time. (KR) Recovery๊ฐ€ ๋ฐ˜๋“œ์‹œ ์ „์ฒด ์žฌ๊ฐ€๋™๊ณผ ์ „์ฒด ์ค‘๋‹จ ์ค‘ ํ•˜๋‚˜๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐฉ์‹์ผ ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. ๋จผ์ € ๊ธˆ์œต Transaction์„ ์‹คํ–‰ํ•  ๊ถŒํ•œ ์—†์ด Read-only Operation๊ณผ Market Monitoring๋งŒ ์žฌ๊ฐœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ๊ณ„์‚ฐ ๊ฒฐ๊ณผ๊ฐ€ ๊ฒ€์ฆ๋˜๋ฉด ๊ฐ•ํ™”๋œ ํ•œ๋„ ๋‚ด์—์„œ ์ œํ•œ์ ์ธ Transaction ๊ธฐ๋Šฅ์„ ๋ณต๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ๊ฒ€์ฆ๊นŒ์ง€ ์™„๋ฃŒํ•œ ์ดํ›„์— ์ „์ฒด Execution Authority๋ฅผ ๋ณต์›ํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„์  ๋ณต๊ตฌ๋Š” ๊ธฐ์กด ์‚ฌ๊ณ ๊ฐ€ ์ฆ‰์‹œ ๋ฐ˜๋ณต๋  ๊ฐ€๋Šฅ์„ฑ์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Recovery๋Š” ๊ฒฝ๊ณผ ์‹œ๊ฐ„์ด ์•„๋‹ˆ๋ผ ๊ฒ€์ฆ๋œ ์•ˆ์ „ ์ˆ˜์ค€์— ๋”ฐ๋ผ ๊ถŒํ•œ์„ ๋ณต๊ตฌํ•˜๋Š” ๊ณผ์ •์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 6/ Rialo and Conditional Recovery (EN) Rialo's Reactive Transactions offer a foundation for exploring recovery workflows driven by predefined conditions. An application could monitor whether oracle consistency, state validation, and authorization requirements have been satisfied. When those conditions are met, a reactive workflow could initiate a controlled recovery procedure. However, reactive execution alone does not guarantee that an agent's internal state is safe or that every external permission has been revoked. Those checks must be implemented through application-specific logic and authorization controls. (KR) Rialo์˜ Reactive Transactions๋Š” ์‚ฌ์ „์— ์ •์˜๋œ ์กฐ๊ฑด์— ๋”ฐ๋ผ Recovery Workflow๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋ฐ˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Application์€ Oracle Consistency, State Validation, Authorization Requirement ๋“ฑ์ด ์ถฉ์กฑ๋๋Š”์ง€ ํ™•์ธํ•˜๋„๋ก ์„ค๊ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์กฐ๊ฑด์ด ๋งŒ์กฑ๋˜๋ฉด Reactive Workflow๋ฅผ ํ†ตํ•ด ํ†ต์ œ๋œ ๋ณต๊ตฌ ์ ˆ์ฐจ๋ฅผ ์‹œ์ž‘ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹ค๋งŒ Reactive Execution ์ž์ฒด๊ฐ€ Agent์˜ Internal State๊ฐ€ ์•ˆ์ „ํ•˜๋‹ค๋Š” ์‚ฌ์‹ค์ด๋‚˜ ๋ชจ๋“  ์™ธ๋ถ€ ๊ถŒํ•œ์ด ์ •์ƒ์ ์œผ๋กœ ์ฒ ํšŒ๋๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ณด์žฅํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒ€์ฆ์€ Application๋ณ„ Logic๊ณผ Authorization Control์„ ํ†ตํ•ด ๊ตฌํ˜„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a demonstration where an autonomous lending agent is suspended after detecting manipulated oracle data. Once the oracle is corrected, the agent attempts to resume execution. The system initially rejects the request because an outdated liquidation instruction remains in its queue. After the instruction is invalidated and risk calculations are refreshed, the agent receives permission to resume monitoring. Limited execution is then restored, followed by full authorization after additional verification. This would demonstrate that recovery is a controlled process rather than a simple restart command. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์กฐ์ž‘๋œ Oracle ๋ฐ์ดํ„ฐ๊ฐ€ ๊ฐ์ง€๋ผ ์ค‘๋‹จ๋œ Autonomous Lending Agent์˜ ๋ณต๊ตฌ Demo์ž…๋‹ˆ๋‹ค. Oracle์ด ์ •์ƒํ™”๋œ ์ดํ›„ Agent๊ฐ€ ์‹คํ–‰ ์žฌ๊ฐœ๋ฅผ ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Queue์— ์˜ค๋ž˜๋œ Liquidation Instruction์ด ๋‚จ์•„ ์žˆ์–ด ์ฒ˜์Œ์—๋Š” ์š”์ฒญ์ด ๊ฑฐ๋ถ€๋ฉ๋‹ˆ๋‹ค. ํ•ด๋‹น ๋ช…๋ น์„ ๋ฌดํšจํ™”ํ•˜๊ณ  Risk Calculation์„ ๋‹ค์‹œ ์ˆ˜ํ–‰ํ•œ ์ดํ›„ Monitoring ๊ถŒํ•œ๋ถ€ํ„ฐ ๋ณต๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด์–ด์„œ ์ œํ•œ์ ์ธ Execution์„ ํ—ˆ์šฉํ•˜๊ณ  ์ถ”๊ฐ€ ๊ฒ€์ฆ์ด ์™„๋ฃŒ๋˜๋ฉด ์ „์ฒด ๊ถŒํ•œ์„ ๋ณต์›ํ•ฉ๋‹ˆ๋‹ค. Recovery๊ฐ€ ๋‹จ์ˆœํ•œ ์žฌ์‹œ์ž‘ ๋ช…๋ น์ด ์•„๋‹ˆ๋ผ ๊ฒ€์ฆ๊ณผ ์Šน์ธ์— ๋”ฐ๋ผ ๋‹จ๊ณ„์ ์œผ๋กœ ์ง„ํ–‰๋˜๋Š” ๊ณผ์ •์ž„์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. (EN) Autonomous finance needs systems that can stop dangerous behavior without permanently sacrificing useful automation. But restarting an agent before its state and permissions are verified could recreate the same failure that caused the shutdown. A reliable recovery process should establish what changed, what was corrected, and which capabilities can safely return. The ability to restart an agent is not proof that it is ready to operate. Safe recovery requires evidence, authorization, and controlled execution. (KR) Autonomous Finance์—๋Š” ์œ„ํ—˜ํ•œ ํ–‰๋™์„ ์ค‘๋‹จํ•˜๋ฉด์„œ๋„ ์ž๋™ํ™”์˜ ์žฅ์ ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Agent์˜ ์ƒํƒœ์™€ ๊ถŒํ•œ์„ ๊ฒ€์ฆํ•˜์ง€ ์•Š์€ ์ฑ„ ์žฌ๊ฐ€๋™ํ•œ๋‹ค๋ฉด ๊ธฐ์กด ์‚ฌ๊ณ ๊ฐ€ ๋ฐ˜๋ณต๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” Recovery Process๋ผ๋ฉด ๋ฌด์—‡์ด ๋ณ€๊ฒฝ๋๊ณ , ์–ด๋–ค ๋ฌธ์ œ๊ฐ€ ์ˆ˜์ •๋์œผ๋ฉฐ, ์–ด๋–ค ๊ธฐ๋Šฅ๋ถ€ํ„ฐ ์•ˆ์ „ํ•˜๊ฒŒ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Agent๋ฅผ ๋‹ค์‹œ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ์•ˆ์ „ํ•˜๊ฒŒ ์‹คํ–‰ํ•  ์ค€๋น„๊ฐ€ ๋๋‹ค๋Š” ์˜๋ฏธ๋Š” ์•„๋‹™๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ Autonomous Finance์—์„œ ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋‹จ์ˆœํ•œ ์ž๋™ ์‹คํ–‰์ด๋‚˜ ์ž๋™ ์ค‘๋‹จ์„ ๋„˜์–ด, ๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ๊ทผ๊ฑฐ์™€ ์Šน์ธ ์ ˆ์ฐจ๋ฅผ ํ†ตํ•ด ์•ˆ์ „ํ•˜๊ฒŒ ์šด์˜์„ ์žฌ๊ฐœํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ๊ฐ€ ์•„๋‹๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
1
2
193
KB์ฆ๊ถŒ๊ณผ ์ˆ˜ํ˜ธ์•„์ด์˜ค๊ฐ€ ์ฑ„๊ถŒ ํ† ํฐ์ฆ๊ถŒ์„ ๋ฐœํ–‰ํ•˜๊ณ  ์ •์‚ฐํ•œ ๋’ค ๋งŒ๊ธฐ ์ƒํ™˜๊นŒ์ง€ ๋๋‚ด๋Š” ์ „ ๊ณผ์ •์„ ์‹ค์ฆํ–ˆ๋‹ค๊ณ  ์˜ค๋Š˜ ๋ฐํ˜”๋‹ค. ๊ณผ๊ธฐ์ •ํ†ต๋ถ€์™€ ํ•œ๊ตญ์ธํ„ฐ๋„ท์ง„ํฅ์›์˜ ๋ธ”๋ก์ฒด์ธ ์‘์šฉ์„œ๋น„์Šค ์ง€์›์‚ฌ์—… ๊ณผ์ œ๋กœ ์ง„ํ–‰๋๋‹ค. ๋Œ€์ƒ์€ ์ผ๋ฐ˜ ๊ธฐ์—…์ด ๋ฐœํ–‰ํ•˜๋Š” ์ „์ž๋‹จ๊ธฐ์‚ฌ์ฑ„, ํ”ํžˆ ๋งํ•˜๋Š” ์ „๋‹จ์ฑ„๋‹ค. ๋ˆˆ์— ๋„๋Š” ๊ฑด ํˆฌ์ž ๋ฐฉ์‹์ด๋‹ค. KB ์นด๋“œ ํฌ์ธํŠธ์ธ ํฌ์ธํŠธ๋ฆฌ๋กœ ์ „๋‹จ์ฑ„ ํ† ํฐ์ฆ๊ถŒ์— ์†Œ์•ก ํˆฌ์žํ•˜๋„๋ก ์„ค๊ณ„ํ–ˆ๋‹ค. ๋ฐœํ–‰ ๊ทœ๊ฒฉ์€ ํ—ˆ๊ฐ€๋œ ์ง€๊ฐ‘๋ผ๋ฆฌ๋งŒ ํ† ํฐ์„ ์˜ฎ๊ธธ ์ˆ˜ ์žˆ๊ฒŒ ๋งŒ๋“  ERC-3643์ด๊ณ  KB์ฆ๊ถŒ์˜ ์ฑ„๊ถŒ ๋ฐ์ดํ„ฐ๋ฅผ API๋กœ ๋ฐ›์•„ ํ† ํฐ์„ ์ฐ์—ˆ๋‹ค. ์ •์‚ฐ์€ ์ฆ๊ถŒ๊ณผ ๋Œ€๊ธˆ์„ ํ•˜๋‚˜์˜ ๊ฑฐ๋ž˜๋กœ ๋ฌถ์–ด ํ•œ์ชฝ์ด ์‹คํŒจํ•˜๋ฉด ์ „์ฒด๋ฅผ ๋˜๋Œ๋ฆฌ๋Š” ๋™์‹œ๊ฒฐ์ œ ๋ฐฉ์‹์ด๊ณ  ๋ธ”๋ก์ฒด์ธ ์›์žฅ๊ณผ ์ฆ๊ถŒ์‚ฌ ๋‚ด๋ถ€ ์›์žฅ์„ ์ž๋™์œผ๋กœ ๋Œ€์กฐํ•˜๋Š” ์žฅ์น˜๋„ ๋ถ™์—ฌ์„œ ์ •ํ•ฉ๋ฅ ์„ 99.9% ์ด์ƒ์œผ๋กœ ๋งž์ท„๋‹ค. ๋งŒ๊ธฐ๊ฐ€ ๋˜๋ฉด ์ƒํ™˜๊ธˆ ์ง€๊ธ‰๊ณผ ํ† ํฐ ์†Œ๊ฐ๊นŒ์ง€ ์‚ฌ๋žŒ ์† ์—†์ด ์ฒ˜๋ฆฌ๋œ๋‹ค. ๋‹ค๋งŒ ์ด๋ฒˆ์— ๋ฐœํ–‰ํ•œ ๊ฑด ๊ฐ€์ƒ์˜ ์ฑ„๊ถŒ์ด๊ณ  ์‹ค์ œ๋กœ ํŒ”๋ฆฐ ์ƒํ’ˆ์€ ์•„๋‹ˆ๋‹ค. KB์ฆ๊ถŒ๋„ ์ œํ•œ๋œ ๋ฒ”์œ„์—์„œ ๊ฐ€์ƒ์˜ ํ† ํฐ์ฆ๊ถŒ์„ ํฌ์ธํŠธ๋กœ ์‚ฌ๋Š” ๊ณผ์ •์„ ๊ฒ€์ฆํ•œ ๋ฐ ์˜์˜๊ฐ€ ์žˆ๋‹ค๊ณ  ํ–ˆ๋‹ค. ์‹ค์ œ ์ƒํ’ˆ์€ ์•„๋‹ˆ์–ด๋„ ์ด๋ฒˆ ์‹ค์ฆ์ด ๋‚˜์˜จ ์‹œ์ ์€ ์˜๋ฏธ๊ฐ€ ์žˆ๋‹ค๊ณ  ๋ณธ๋‹ค. ํ† ํฐ์ฆ๊ถŒ์„ ์ œ๋„๊ถŒ์— ๋“ค์ด๋Š” ๊ฐœ์ • ์ „์ž์ฆ๊ถŒ๋ฒ•๊ณผ ์ž๋ณธ์‹œ์žฅ๋ฒ•์ด 2027๋…„ 2์›” 4์ผ ์‹œํ–‰๋˜๊ณ  ๊ธˆ์œต์œ„๋Š” 9์›” ์ •์ฑ…๋ฐฉํ–ฅ์—์„œ ์‚ฌ๋ชจ์ฑ„๊ถŒ๊ณผ ์‚ฌ๋ชจMMF ๊ฐ™์€ ์ •ํ˜•์ฆ๊ถŒ๋„ ๋ฐœํ–‰ ๋Œ€์ƒ์— ๋„ฃ์—ˆ๋‹ค. ์ฆ๊ถŒ์‚ฌ๊ฐ€ ์‹œํ–‰์„ ์„ ๋‹ฌ ๊ฐ€๊นŒ์ด ์•ž๋‘๊ณ  ์ž๊ธฐ ์‹œ์Šคํ…œ์— ๋ถ™์—ฌ ๋๊นŒ์ง€ ๋Œ๋ ค ๋ณธ ์…ˆ์ด๋ผ ํฌ์ธํŠธ๋กœ ์ฑ„๊ถŒ ์กฐ๊ฐ์„ ์‚ฌ๋Š” ๊ตฌ์กฐ๊ฐ€ ์‹ค์ œ ์ƒํ’ˆ ํ›„๋ณด๋กœ ์˜ฌ๋ผ์˜ฌ์ง€ ์ง€์ผœ๋ณผ ๋งŒํ•˜๋‹ค. KB์ฆ๊ถŒ์€ ์Šคํ…Œ์ด๋ธ”์ฝ”์ธ ๋ฒ•์ œํ™”๊ฐ€ ๋๋‚˜๋ฉด KB๊ธˆ์œต ์Šคํ…Œ์ด๋ธ”์ฝ”์ธ๊ณผ ์—ฐ๊ณ„ํ•˜๋Š” ๊ฒƒ๋„ ์ด๋ฒˆ์— ์‹œํ—˜ํ•ด ๋ณผ ์ˆ˜ ์žˆ์—ˆ๋‹ค๊ณ  ํ–ˆ๋Š”๋ฐ ๊ทธ ๋ถ€๋ถ„์€ ์•„์ง ๊ณ„ํš ์ˆ˜์ค€์ด๋‹ค.
37
๐Ÿ“š Rialo Research Series 65: Agent Recovery Authorization - When Is It Safe to Restart an Autonomous Agent? (Kor : ์—์ด์ „ํŠธ ๋ณต๊ตฌ ์Šน์ธ - ์ค‘๋‹จ๋œ ์ž์œจ Agent๋ฅผ ์–ธ์ œ ๋‹ค์‹œ ์‹คํ–‰ํ•ด๋„ ์•ˆ์ „ํ• ๊นŒ?) 1/ Intro: Stopping an Agent Is Only Half the Solution (EN) In the previous chapter, we explored how emergency shutdown authority could prevent an autonomous agent from causing further financial damage. But suspending execution creates another problem. Imagine a lending agent that was stopped after detecting inconsistent oracle prices. The oracle eventually returns to normal, but should the agent immediately resume its operations? What if its pending instructions, permissions, or internal state still reflect the corrupted data? Stopping an unsafe agent is one challenge. Proving that it is safe to restart is another. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” Emergency Shutdown Authority๋ฅผ ํ†ตํ•ด ์œ„ํ—˜ํ•œ Agent์˜ ์ถ”๊ฐ€์ ์ธ ๊ธˆ์œต ํ”ผํ•ด๋ฅผ ๋ง‰๋Š” ๋ฐฉ๋ฒ•์„ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์‹คํ–‰์„ ์ค‘๋‹จํ•˜๋ฉด ๋˜ ๋‹ค๋ฅธ ๋ฌธ์ œ๊ฐ€ ์ƒ๊น๋‹ˆ๋‹ค. Oracle ๊ฐ€๊ฒฉ ๋ถˆ์ผ์น˜๊ฐ€ ๊ฐ์ง€๋ผ Lending Agent๊ฐ€ ์ค‘๋‹จ๋๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ดํ›„ Oracle์ด ์ •์ƒํ™”๋๋‹ค๊ณ  ํ•ด์„œ Agent๋ฅผ ์ฆ‰์‹œ ์žฌ๊ฐ€๋™ํ•ด๋„ ๊ดœ์ฐฎ์„๊นŒ์š”? ๊ธฐ์กด Pending Instruction์ด๋‚˜ Permission, Internal State์— ์—ฌ์ „ํžˆ ์ž˜๋ชป๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ฐ˜์˜๋ผ ์žˆ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„ํ—˜ํ•œ Agent๋ฅผ ์ค‘๋‹จํ•˜๋Š” ๊ฒƒ๊ณผ ์•ˆ์ „ํ•˜๊ฒŒ ๋‹ค์‹œ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Œ์„ ์ฆ๋ช…ํ•˜๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. 2/ Why Normal Data Doesn't Mean Normal Operation (EN) Suppose an agent calculates liquidation risk using a corrupted collateral price. The oracle is corrected, but the agent has already stored an incorrect risk assessment. If execution resumes without refreshing that assessment, the agent could submit a liquidation based on outdated information. Restoring the original data source does not automatically repair every downstream decision. Recovery must verify both the external inputs and the internal state derived from them. (KR) Agent๊ฐ€ ์ž˜๋ชป๋œ ๋‹ด๋ณด ๊ฐ€๊ฒฉ์„ ๊ธฐ๋ฐ˜์œผ๋กœ Liquidation Risk๋ฅผ ๊ณ„์‚ฐํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. Oracle ๊ฐ€๊ฒฉ์€ ์ˆ˜์ •๋์ง€๋งŒ Agent ๋‚ด๋ถ€์—๋Š” ์ด๋ฏธ ์ž˜๋ชป๋œ Risk Assessment๊ฐ€ ์ €์žฅ๋ผ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ์ƒํƒœ์—์„œ ์‹คํ–‰์„ ์žฌ๊ฐœํ•˜๋ฉด ์˜ค๋ž˜๋œ ํŒ๋‹จ์„ ๊ธฐ๋ฐ˜์œผ๋กœ Liquidation Transaction์„ ์ œ์ถœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์™ธ๋ถ€ ๋ฐ์ดํ„ฐ๊ฐ€ ์ •์ƒํ™”๋๋‹ค๊ณ  ํ•ด์„œ ๊ทธ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ƒ์„ฑ๋œ ๋ชจ๋“  ํŒ๋‹จ๊นŒ์ง€ ๋ณต๊ตฌ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. Recovery ๊ณผ์ •์—์„œ๋Š” External Input๊ณผ Internal State๋ฅผ ํ•จ๊ป˜ ๊ฒ€์ฆํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 3/ Recovery Preconditions (EN) An agent should not restart simply because a timeout has expired or an administrator presses a button. Applications could define explicit recovery conditions. For example, oracle prices must remain consistent across multiple observations, affected risk calculations must be refreshed, and unsafe pending instructions must be invalidated. Previously granted permissions should also be reviewed. Only after these conditions are satisfied should the system consider restoring execution authority. (KR) ์ผ์ • ์‹œ๊ฐ„์ด ์ง€๋‚ฌ๊ฑฐ๋‚˜ ๊ด€๋ฆฌ์ž๊ฐ€ ๋ฒ„ํŠผ์„ ๋ˆŒ๋ €๋‹ค๋Š” ์ด์œ ๋งŒ์œผ๋กœ Agent๋ฅผ ์žฌ๊ฐ€๋™ํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค. Application์€ ๋ช…ํ™•ํ•œ Recovery Condition์„ ์ •์˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์—ฌ๋Ÿฌ ์ฐจ๋ก€ ๊ด€์ธก์—์„œ Oracle ๊ฐ€๊ฒฉ์ด ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•ด์•ผ ํ•˜๊ณ , ์˜ํ–ฅ์„ ๋ฐ›์€ Risk Calculation์„ ๋‹ค์‹œ ์ˆ˜ํ–‰ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์œ„ํ—˜ํ•œ Pending Instruction์€ ๋ฌดํšจํ™”ํ•˜๊ณ  ๊ธฐ์กด Permission๋„ ์žฌ๊ฒ€ํ† ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์กฐ๊ฑด์ด ์ถฉ์กฑ๋œ ์ดํ›„์— ์‹คํ–‰ ๊ถŒํ•œ ๋ณต๊ตฌ๋ฅผ ๊ฒ€ํ† ํ•˜๋Š” ๊ตฌ์กฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 4/ Who Approves the Restart? (EN) The authority to suspend an agent does not necessarily need to be identical to the authority to restart it. Emergency intervention prioritizes speed. Recovery authorization prioritizes verification. A security operator might be allowed to suspend execution immediately, while restarting high-risk operations requires additional approval. For example, a multi-signature authorization process could prevent one compromised administrator from restoring unsafe permissions. However, the approval structure must remain practical enough to avoid unnecessary service disruption. (KR) Agent๋ฅผ ์ค‘๋‹จํ•  ์ˆ˜ ์žˆ๋Š” ๊ถŒํ•œ๊ณผ ๋‹ค์‹œ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋Š” ๊ถŒํ•œ์ด ๋ฐ˜๋“œ์‹œ ๊ฐ™์„ ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. Emergency Intervention์—์„œ๋Š” ๋Œ€์‘ ์†๋„๊ฐ€ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด Recovery Authorization์—์„œ๋Š” ๊ฒ€์ฆ ๊ณผ์ •์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์šด์˜์ž๊ฐ€ ์ฆ‰์‹œ ์‹คํ–‰์„ ์ค‘๋‹จํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋˜, ๊ณ ์œ„ํ—˜ ๊ธฐ๋Šฅ์˜ ์žฌ๊ฐ€๋™์—๋Š” ์ถ”๊ฐ€ ์Šน์ธ์„ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Multi-signature Authorization์„ ํ™œ์šฉํ•˜๋ฉด ํƒˆ์ทจ๋œ ๊ด€๋ฆฌ์ž ํ•œ ๋ช…์ด ์œ„ํ—˜ํ•œ ๊ถŒํ•œ์„ ์ž„์˜๋กœ ๋ณต๊ตฌํ•˜๋Š” ์ƒํ™ฉ์„ ๋ฐฉ์ง€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋‹ค๋งŒ ์Šน์ธ ์ ˆ์ฐจ๊ฐ€ ์ง€๋‚˜์น˜๊ฒŒ ๋ณต์žกํ•ด ์ •์ƒ์ ์ธ ์„œ๋น„์Šค ๋ณต๊ตฌ๊นŒ์ง€ ๋ฐฉํ•ดํ•˜์ง€ ์•Š๋„๋ก ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 5/ Gradual Recovery Instead of Full Restart (EN) Recovery does not have to be an all-or-nothing decision. An agent could first resume read-only operations and market monitoring without permission to execute financial transactions. After its calculations are verified, limited transaction capabilities could be restored under stricter limits. Full execution authority would return only after additional checks. This staged approach reduces the risk of immediately repeating the original failure. Recovery should restore capabilities according to verified safety, not merely elapsed time. (KR) Recovery๊ฐ€ ๋ฐ˜๋“œ์‹œ ์ „์ฒด ์žฌ๊ฐ€๋™๊ณผ ์ „์ฒด ์ค‘๋‹จ ์ค‘ ํ•˜๋‚˜๋ฅผ ์„ ํƒํ•˜๋Š” ๋ฐฉ์‹์ผ ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. ๋จผ์ € ๊ธˆ์œต Transaction์„ ์‹คํ–‰ํ•  ๊ถŒํ•œ ์—†์ด Read-only Operation๊ณผ Market Monitoring๋งŒ ์žฌ๊ฐœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ๊ณ„์‚ฐ ๊ฒฐ๊ณผ๊ฐ€ ๊ฒ€์ฆ๋˜๋ฉด ๊ฐ•ํ™”๋œ ํ•œ๋„ ๋‚ด์—์„œ ์ œํ•œ์ ์ธ Transaction ๊ธฐ๋Šฅ์„ ๋ณต๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ถ”๊ฐ€ ๊ฒ€์ฆ๊นŒ์ง€ ์™„๋ฃŒํ•œ ์ดํ›„์— ์ „์ฒด Execution Authority๋ฅผ ๋ณต์›ํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋‹จ๊ณ„์  ๋ณต๊ตฌ๋Š” ๊ธฐ์กด ์‚ฌ๊ณ ๊ฐ€ ์ฆ‰์‹œ ๋ฐ˜๋ณต๋  ๊ฐ€๋Šฅ์„ฑ์„ ์ค„์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Recovery๋Š” ๊ฒฝ๊ณผ ์‹œ๊ฐ„์ด ์•„๋‹ˆ๋ผ ๊ฒ€์ฆ๋œ ์•ˆ์ „ ์ˆ˜์ค€์— ๋”ฐ๋ผ ๊ถŒํ•œ์„ ๋ณต๊ตฌํ•˜๋Š” ๊ณผ์ •์ด์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 6/ Rialo and Conditional Recovery (EN) Rialo's Reactive Transactions offer a foundation for exploring recovery workflows driven by predefined conditions. An application could monitor whether oracle consistency, state validation, and authorization requirements have been satisfied. When those conditions are met, a reactive workflow could initiate a controlled recovery procedure. However, reactive execution alone does not guarantee that an agent's internal state is safe or that every external permission has been revoked. Those checks must be implemented through application-specific logic and authorization controls. (KR) Rialo์˜ Reactive Transactions๋Š” ์‚ฌ์ „์— ์ •์˜๋œ ์กฐ๊ฑด์— ๋”ฐ๋ผ Recovery Workflow๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๋ฐ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋ฐ˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Application์€ Oracle Consistency, State Validation, Authorization Requirement ๋“ฑ์ด ์ถฉ์กฑ๋๋Š”์ง€ ํ™•์ธํ•˜๋„๋ก ์„ค๊ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์กฐ๊ฑด์ด ๋งŒ์กฑ๋˜๋ฉด Reactive Workflow๋ฅผ ํ†ตํ•ด ํ†ต์ œ๋œ ๋ณต๊ตฌ ์ ˆ์ฐจ๋ฅผ ์‹œ์ž‘ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹ค๋งŒ Reactive Execution ์ž์ฒด๊ฐ€ Agent์˜ Internal State๊ฐ€ ์•ˆ์ „ํ•˜๋‹ค๋Š” ์‚ฌ์‹ค์ด๋‚˜ ๋ชจ๋“  ์™ธ๋ถ€ ๊ถŒํ•œ์ด ์ •์ƒ์ ์œผ๋กœ ์ฒ ํšŒ๋๋‹ค๋Š” ์‚ฌ์‹ค์„ ๋ณด์žฅํ•˜์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฒ€์ฆ์€ Application๋ณ„ Logic๊ณผ Authorization Control์„ ํ†ตํ•ด ๊ตฌํ˜„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a demonstration where an autonomous lending agent is suspended after detecting manipulated oracle data. Once the oracle is corrected, the agent attempts to resume execution. The system initially rejects the request because an outdated liquidation instruction remains in its queue. After the instruction is invalidated and risk calculations are refreshed, the agent receives permission to resume monitoring. Limited execution is then restored, followed by full authorization after additional verification. This would demonstrate that recovery is a controlled process rather than a simple restart command. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์กฐ์ž‘๋œ Oracle ๋ฐ์ดํ„ฐ๊ฐ€ ๊ฐ์ง€๋ผ ์ค‘๋‹จ๋œ Autonomous Lending Agent์˜ ๋ณต๊ตฌ Demo์ž…๋‹ˆ๋‹ค. Oracle์ด ์ •์ƒํ™”๋œ ์ดํ›„ Agent๊ฐ€ ์‹คํ–‰ ์žฌ๊ฐœ๋ฅผ ์š”์ฒญํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Queue์— ์˜ค๋ž˜๋œ Liquidation Instruction์ด ๋‚จ์•„ ์žˆ์–ด ์ฒ˜์Œ์—๋Š” ์š”์ฒญ์ด ๊ฑฐ๋ถ€๋ฉ๋‹ˆ๋‹ค. ํ•ด๋‹น ๋ช…๋ น์„ ๋ฌดํšจํ™”ํ•˜๊ณ  Risk Calculation์„ ๋‹ค์‹œ ์ˆ˜ํ–‰ํ•œ ์ดํ›„ Monitoring ๊ถŒํ•œ๋ถ€ํ„ฐ ๋ณต๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์ด์–ด์„œ ์ œํ•œ์ ์ธ Execution์„ ํ—ˆ์šฉํ•˜๊ณ  ์ถ”๊ฐ€ ๊ฒ€์ฆ์ด ์™„๋ฃŒ๋˜๋ฉด ์ „์ฒด ๊ถŒํ•œ์„ ๋ณต์›ํ•ฉ๋‹ˆ๋‹ค. Recovery๊ฐ€ ๋‹จ์ˆœํ•œ ์žฌ์‹œ์ž‘ ๋ช…๋ น์ด ์•„๋‹ˆ๋ผ ๊ฒ€์ฆ๊ณผ ์Šน์ธ์— ๋”ฐ๋ผ ๋‹จ๊ณ„์ ์œผ๋กœ ์ง„ํ–‰๋˜๋Š” ๊ณผ์ •์ž„์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. (EN) Autonomous finance needs systems that can stop dangerous behavior without permanently sacrificing useful automation. But restarting an agent before its state and permissions are verified could recreate the same failure that caused the shutdown. A reliable recovery process should establish what changed, what was corrected, and which capabilities can safely return. The ability to restart an agent is not proof that it is ready to operate. Safe recovery requires evidence, authorization, and controlled execution. (KR) Autonomous Finance์—๋Š” ์œ„ํ—˜ํ•œ ํ–‰๋™์„ ์ค‘๋‹จํ•˜๋ฉด์„œ๋„ ์ž๋™ํ™”์˜ ์žฅ์ ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Agent์˜ ์ƒํƒœ์™€ ๊ถŒํ•œ์„ ๊ฒ€์ฆํ•˜์ง€ ์•Š์€ ์ฑ„ ์žฌ๊ฐ€๋™ํ•œ๋‹ค๋ฉด ๊ธฐ์กด ์‚ฌ๊ณ ๊ฐ€ ๋ฐ˜๋ณต๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” Recovery Process๋ผ๋ฉด ๋ฌด์—‡์ด ๋ณ€๊ฒฝ๋๊ณ , ์–ด๋–ค ๋ฌธ์ œ๊ฐ€ ์ˆ˜์ •๋์œผ๋ฉฐ, ์–ด๋–ค ๊ธฐ๋Šฅ๋ถ€ํ„ฐ ์•ˆ์ „ํ•˜๊ฒŒ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Agent๋ฅผ ๋‹ค์‹œ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์ด ์•ˆ์ „ํ•˜๊ฒŒ ์‹คํ–‰ํ•  ์ค€๋น„๊ฐ€ ๋๋‹ค๋Š” ์˜๋ฏธ๋Š” ์•„๋‹™๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ Autonomous Finance์—์„œ ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋‹จ์ˆœํ•œ ์ž๋™ ์‹คํ–‰์ด๋‚˜ ์ž๋™ ์ค‘๋‹จ์„ ๋„˜์–ด, ๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ๊ทผ๊ฑฐ์™€ ์Šน์ธ ์ ˆ์ฐจ๋ฅผ ํ†ตํ•ด ์•ˆ์ „ํ•˜๊ฒŒ ์šด์˜์„ ์žฌ๊ฐœํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ๊ฐ€ ์•„๋‹๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
๐Ÿ“š Rialo Research Series 64: Agent Shutdown Authority - Who Can Stop an Autonomous Agent? (Kor : ์—์ด์ „ํŠธ ์ค‘๋‹จ ๊ถŒํ•œ - ์ž์œจ์ ์œผ๋กœ ์›€์ง์ด๋Š” Agent๋ฅผ ๋ˆ„๊ฐ€ ๋ฉˆ์ถœ ์ˆ˜ ์žˆ์„๊นŒ?) 1/ Intro: Accountability Is Not Enough (EN) In the previous chapter, we explored how responsibility could be traced across multiple agents involved in a failed financial decision. But identifying responsibility after an incident doesn't prevent further damage while the incident is unfolding. Imagine an autonomous trading agent repeatedly executing transactions based on corrupted market data. Investigators may eventually determine what went wrong. Meanwhile, the agent continues trading. Accountability explains a failure. Shutdown authority determines whether the failure can be contained. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ์—ฌ๋Ÿฌ Agent๊ฐ€ ๊ด€์—ฌํ•œ ์ž˜๋ชป๋œ ๊ธˆ์œต ์˜์‚ฌ๊ฒฐ์ •์—์„œ ์ฑ…์ž„์„ ์–ด๋–ป๊ฒŒ ์ถ”์ ํ•  ์ˆ˜ ์žˆ์„์ง€ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์‚ฌ๊ณ  ์ดํ›„ ์ฑ…์ž„์„ ํ™•์ธํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ๋Š” ์ง„ํ–‰ ์ค‘์ธ ํ”ผํ•ด๋ฅผ ๋ง‰์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. Autonomous Trading Agent๊ฐ€ ์˜ค์—ผ๋œ Market Data๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ณ„์† ๊ฑฐ๋ž˜ํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋‚˜์ค‘์— ์‚ฌ๊ณ  ์›์ธ์„ ๋ฐํ˜€๋‚ด๋”๋ผ๋„ ๊ทธ๋™์•ˆ Agent๋Š” ๊ฑฐ๋ž˜๋ฅผ ๊ณ„์† ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Accountability๊ฐ€ ์‹คํŒจ๋ฅผ ์„ค๋ช…ํ•˜๋Š” ๊ณผ์ •์ด๋ผ๋ฉด Shutdown Authority๋Š” ํ”ผํ•ด ํ™•์‚ฐ์„ ๋ง‰๋Š” ํ†ต์ œ ์ˆ˜๋‹จ์ž…๋‹ˆ๋‹ค. 2/ The Problem With Autonomous Execution (EN) Traditional financial systems generally have operators who can suspend services or restrict transactions. Autonomous agents complicate this arrangement. An agent may operate continuously, interact with multiple protocols, and trigger transactions without requesting approval for every action. Disabling its user interface may not stop already authorized workflows or pending transactions. Effective shutdown mechanisms must therefore operate at the level where execution authority is actually enforced. (KR) ์ „ํ†ต์ ์ธ ๊ธˆ์œต ์‹œ์Šคํ…œ์—์„œ๋Š” ์šด์˜์ž๊ฐ€ ์„œ๋น„์Šค๋ฅผ ์ค‘๋‹จํ•˜๊ฑฐ๋‚˜ ๊ฑฐ๋ž˜๋ฅผ ์ œํ•œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Autonomous Agent๋Š” ์ƒํ™ฉ์ด ์กฐ๊ธˆ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. Agent๋Š” ์ง€์†์ ์œผ๋กœ ์ž‘๋™ํ•˜๋ฉด์„œ ์—ฌ๋Ÿฌ Protocol๊ณผ ์ƒํ˜ธ์ž‘์šฉํ•˜๊ณ , ๋งค๋ฒˆ ์Šน์ธ์„ ๋ฐ›์ง€ ์•Š๊ณ  Transaction์„ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹จ์ˆœํžˆ ์‚ฌ์šฉ์ž ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ๋น„ํ™œ์„ฑํ™”ํ•˜๋”๋ผ๋„ ์ด๋ฏธ ์Šน์ธ๋œ Workflow๋‚˜ Pending Transaction๊นŒ์ง€ ์ค‘๋‹จ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹ค์ œ ์‹คํ–‰ ๊ถŒํ•œ์ด ์ ์šฉ๋˜๋Š” ์ง€์ ์—์„œ ํ†ต์ œํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 3/ Three Levels of Intervention (EN) Not every incident requires permanently disabling an agent. A system could distinguish between three intervention levels. The first restricts specific actions, such as borrowing or liquidation. The second temporarily suspends the agent's execution permissions while preserving its records and state. The third revokes critical authorization when an agent is compromised or cannot be safely recovered. These controls should have clearly defined triggers and authorization requirements. (KR) ๋ชจ๋“  ์‚ฌ๊ณ ์—์„œ Agent๋ฅผ ์˜๊ตฌ์ ์œผ๋กœ ์ค‘๋‹จํ•  ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. ๊ฐœ์ž… ์ˆ˜์ค€์„ ์„ธ ๋‹จ๊ณ„๋กœ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋Š” Borrowing์ด๋‚˜ Liquidation ๊ฐ™์€ ํŠน์ • ํ–‰๋™๋งŒ ์ œํ•œํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ๋Š” Agent์˜ ๊ธฐ๋ก๊ณผ ์ƒํƒœ๋ฅผ ์œ ์ง€ํ•˜๋ฉด์„œ ์‹คํ–‰ ๊ถŒํ•œ์„ ์ผ์‹œ์ ์œผ๋กœ ์ •์ง€ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„ธ ๋ฒˆ์งธ๋Š” Agent๊ฐ€ ํƒˆ์ทจ๋๊ฑฐ๋‚˜ ์•ˆ์ „ํ•˜๊ฒŒ ๋ณต๊ตฌํ•˜๊ธฐ ์–ด๋ ค์šด ๊ฒฝ์šฐ ํ•ต์‹ฌ Authorization์„ ์ฒ ํšŒํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๊ฐ ๋‹จ๊ณ„๋งˆ๋‹ค ๋ฐœ๋™ ์กฐ๊ฑด๊ณผ ์Šน์ธ ๊ถŒํ•œ์ด ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜๋ผ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 4/ Who Holds the Emergency Authority? (EN) Giving one administrator unrestricted shutdown authority creates a centralization risk. But requiring a full governance vote during an active exploit could make intervention too slow. A possible compromise is separating emergency suspension from permanent policy changes. For example, a designated security committee could temporarily pause high-risk operations, while restoring permissions requires additional review. The challenge is balancing response speed with protection against authority abuse. (KR) ํ•œ ๋ช…์˜ ๊ด€๋ฆฌ์ž์—๊ฒŒ ๋ฌด์ œํ•œ์ ์ธ Shutdown Authority๋ฅผ ๋ถ€์—ฌํ•˜๋ฉด ์ค‘์•™ํ™” ์œ„ํ—˜์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ์ง„ํ–‰ ์ค‘์ธ ๊ณต๊ฒฉ์„ ๋ง‰๊ธฐ ์œ„ํ•ด ๋ชจ๋“  ๊ฒฐ์ •์„ Governance Vote๋กœ ์ฒ˜๋ฆฌํ•œ๋‹ค๋ฉด ๋Œ€์‘์ด ์ง€๋‚˜์น˜๊ฒŒ ๋Šฆ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์€ ๊ธด๊ธ‰ ์ค‘๋‹จ ๊ถŒํ•œ๊ณผ ์˜๊ตฌ์ ์ธ ์ •์ฑ… ๋ณ€๊ฒฝ ๊ถŒํ•œ์„ ๋ถ„๋ฆฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์ง€์ •๋œ Security Committee๊ฐ€ ๊ณ ์œ„ํ—˜ ์ž‘์—…์„ ์ผ์‹œ ์ค‘๋‹จํ•˜๋˜, ๊ถŒํ•œ ๋ณต๊ตฌ์—๋Š” ๋ณ„๋„์˜ ๊ฒ€ํ†  ์ ˆ์ฐจ๋ฅผ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•ต์‹ฌ์€ ๋น ๋ฅธ ์‚ฌ๊ณ  ๋Œ€์‘๊ณผ ๊ถŒํ•œ ๋‚จ์šฉ ๋ฐฉ์ง€ ์‚ฌ์ด์˜ ๊ท ํ˜•์ž…๋‹ˆ๋‹ค. 5/ What Happens to Pending Transactions? (EN) Stopping an agent does not necessarily cancel everything it has already initiated. Some transactions may be waiting for inclusion, while others may already be finalized. External protocols may also retain permissions granted before the shutdown. A reliable intervention design must distinguish between preventing new instructions, revoking existing permissions, and handling transactions that can no longer be stopped. A shutdown button is only effective when its actual scope is understood. (KR) Agent๋ฅผ ์ค‘๋‹จํ•œ๋‹ค๊ณ  ํ•ด์„œ ์ด๋ฏธ ์‹œ์ž‘๋œ ๋ชจ๋“  ์ž‘์—…์ด ์ทจ์†Œ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ผ๋ถ€ Transaction์€ ์•„์ง ์ฒ˜๋ฆฌ ๋Œ€๊ธฐ ์ค‘์ผ ์ˆ˜ ์žˆ๊ณ , ๋‹ค๋ฅธ Transaction์€ ์ด๋ฏธ Finalized๋์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์™ธ๋ถ€ Protocol์— ๋ถ€์—ฌํ•œ ๊ถŒํ•œ ์—ญ์‹œ ์ค‘๋‹จ ์ดํ›„์—๋„ ๋‚จ์•„ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ƒˆ๋กœ์šด ๋ช…๋ น ์ฐจ๋‹จ, ๊ธฐ์กด ๊ถŒํ•œ ์ฒ ํšŒ, ์ค‘๋‹จํ•  ์ˆ˜ ์—†๋Š” ๊ฑฐ๋ž˜์˜ ์‚ฌํ›„ ์ฒ˜๋ฆฌ๊นŒ์ง€ ๊ตฌ๋ถ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Shutdown ๊ธฐ๋Šฅ์˜ ์‹คํšจ์„ฑ์€ ๋ฒ„ํŠผ์˜ ์กด์žฌ๊ฐ€ ์•„๋‹ˆ๋ผ ์‹ค์ œ๋กœ ์–ด๋””๊นŒ์ง€ ํ†ต์ œํ•  ์ˆ˜ ์žˆ๋А๋ƒ์— ๋‹ฌ๋ ค ์žˆ์Šต๋‹ˆ๋‹ค. 6/ Rialo and Reactive Safety Controls (EN) Rialo's Reactive Transactions provide an interesting foundation for designing condition-based safety responses. An application could monitor predefined risk conditions and trigger restrictions when those conditions are satisfied. For example, repeated oracle inconsistencies might activate a temporary limitation on liquidation operations. However, reactive execution does not automatically grant universal authority to stop agents or reverse transactions. The application must still implement permission boundaries, emergency roles, and recovery procedures. (KR) Rialo์˜ Reactive Transactions๋Š” ์กฐ๊ฑด ๊ธฐ๋ฐ˜ Safety Control์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋ฐ˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Application์€ ์‚ฌ์ „์— ์ •์˜ํ•œ Risk Condition์„ ๊ฐ์‹œํ•˜๊ณ , ํ•ด๋‹น ์กฐ๊ฑด์ด ์ถฉ์กฑ๋˜๋ฉด ์ œํ•œ ์กฐ์น˜๋ฅผ ์‹คํ–‰ํ•˜๋„๋ก ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Oracle ๋ฐ์ดํ„ฐ ๋ถˆ์ผ์น˜๊ฐ€ ๋ฐ˜๋ณต๋˜๋ฉด Liquidation ๊ธฐ๋Šฅ์„ ์ผ์‹œ์ ์œผ๋กœ ์ œํ•œํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ Reactive Execution์ด ๋ชจ๋“  Agent๋ฅผ ์ค‘๋‹จํ•˜๊ฑฐ๋‚˜ Transaction์„ ๋˜๋Œ๋ฆด ์ˆ˜ ์žˆ๋Š” ๋ณดํŽธ์ ์ธ ๊ถŒํ•œ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์‹ค์ œ Permission Boundary, Emergency Role, Recovery Procedure๋Š” Application์—์„œ ๊ตฌํ˜„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a demonstration involving an autonomous lending agent operating under normal market conditions. During execution, the oracle begins returning inconsistent prices. The system detects repeated anomalies and automatically restricts new liquidation instructions while preserving the agent's decision history. An authorized operator then reviews the incident and decides whether execution can safely resume. The demonstration should also show which pending transactions remain unaffected by the shutdown. That would reveal the practical limits of automated intervention rather than presenting shutdown as an absolute guarantee. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์ •์ƒ์ ์ธ ์‹œ์žฅ ํ™˜๊ฒฝ์—์„œ ์ž‘๋™ํ•˜๋Š” Autonomous Lending Agent Demo์ž…๋‹ˆ๋‹ค. ์‹คํ–‰ ๋„์ค‘ Oracle์ด ์„œ๋กœ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ๊ฐ€๊ฒฉ์„ ๋ฐ˜ํ™˜ํ•˜๊ธฐ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์‹œ์Šคํ…œ์€ ๋ฐ˜๋ณต๋˜๋Š” ์ด์ƒ ์ง•ํ›„๋ฅผ ๊ฐ์ง€ํ•˜๊ณ , Agent์˜ Decision History๋ฅผ ๋ณด์กดํ•˜๋ฉด์„œ ์ƒˆ๋กœ์šด Liquidation ๋ช…๋ น์„ ์ž๋™์œผ๋กœ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„ ๊ถŒํ•œ์„ ๊ฐ€์ง„ ์šด์˜์ž๊ฐ€ ์‚ฌ๊ณ ๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ์‹คํ–‰ ์žฌ๊ฐœ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ๋™์‹œ์— Shutdown ์ดํ›„์—๋„ ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š๋Š” Pending Transaction์ด ๋ฌด์—‡์ธ์ง€ ๋ณด์—ฌ์ฃผ๋ฉด ์ข‹๊ฒ ์Šต๋‹ˆ๋‹ค. ์ž๋™ ์ค‘๋‹จ์ด ๋ชจ๋“  ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•œ๋‹ค๋Š” ์ฃผ์žฅ๋ณด๋‹ค ์‹ค์ œ ํ†ต์ œ ๋ฒ”์œ„์™€ ํ•œ๊ณ„๋ฅผ ๊ฒ€์ฆํ•˜๋Š” ๊ณผ์ •์ด ๋” ์ค‘์š”ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. (EN) Autonomous finance needs more than agents capable of making independent decisions. It also needs mechanisms that can intervene when those decisions become dangerous. The ability to suspend execution, revoke permissions, and recover safely should be designed before an incident occurs, not improvised afterward. The real measure of autonomy may not be how long an agent can operate without humans, but how safely humans can regain control when necessary. (KR) Autonomous Finance์—๋Š” ๋…๋ฆฝ์ ์œผ๋กœ ํŒ๋‹จํ•˜๊ณ  ์‹คํ–‰ํ•˜๋Š” Agent๋งŒ ํ•„์š”ํ•œ ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ ํŒ๋‹จ์ด ์œ„ํ—˜ํ•ด์กŒ์„ ๋•Œ ๊ฐœ์ž…ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ๋„ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์‹คํ–‰ ์ค‘๋‹จ, ๊ถŒํ•œ ์ฒ ํšŒ, ์•ˆ์ „ํ•œ ๋ณต๊ตฌ ์ ˆ์ฐจ๋Š” ์‚ฌ๊ณ ๊ฐ€ ๋ฐœ์ƒํ•œ ๋’ค ๊ณ ๋ฏผํ•  ๋ฌธ์ œ๊ฐ€ ์•„๋‹ˆ๋ผ ์‚ฌ์ „์— ์„ค๊ณ„ํ•ด์•ผ ํ•  ์š”์†Œ์ž…๋‹ˆ๋‹ค. Agent๊ฐ€ ์–ผ๋งˆ๋‚˜ ์˜ค๋žซ๋™์•ˆ ์‚ฌ๋žŒ ์—†์ด ์ž‘๋™ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋„ ์ค‘์š”ํ•˜๊ฒ ์ง€๋งŒ, ์œ„ํ—˜ํ•œ ์ƒํ™ฉ์—์„œ ์‚ฌ๋žŒ์ด ์–ผ๋งˆ๋‚˜ ์•ˆ์ „ํ•˜๊ฒŒ ํ†ต์ œ๊ถŒ์„ ๋˜์ฐพ์„ ์ˆ˜ ์žˆ๋Š”์ง€๋„ ์ค‘์š”ํ•˜์ง€ ์•Š์„๊นŒ์š”? ์ง„์ •ํ•œ ์ž์œจ์„ฑ์€ ํ†ต์ œํ•  ์ˆ˜ ์—†๋Š” ์ž์œ ๊ฐ€ ์•„๋‹ˆ๋ผ ํ•„์š”ํ•  ๋•Œ ์•ˆ์ „ํ•˜๊ฒŒ ๊ฐœ์ž…ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ ์œ„์—์„œ ์™„์„ฑ๋œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
1
152
ํ† ์Šค ์ฝ”์ธ์ง€๊ฐ‘, ์“ฐ๊ธฐ ์ „์— ์•Œ์•„ ๋‘˜ ๋„คํŠธ์›Œํฌยท๊ฐ€์Šค๋น„ยทํ‚ค ๊ด€๋ฆฌ ํ† ์Šค๊ฐ€ ์ด๋ฒˆ ์ฃผ ๋น„์ˆ˜ํƒํ˜• ๊ฐ€์ƒ์ž์‚ฐ ์ง€๊ฐ‘์ธ ํ† ์Šค ์ฝ”์ธ์ง€๊ฐ‘์„ ๋‚ด๋†“๋Š”๋‹ค๊ณ  16์ผ ๋ฐํ˜”๋‹ค. ํ† ์Šค ๊ฐ€์ž…์ž๋ฉด ๋ณ„๋„ ์‹ ์ฒญ ์—†์ด ์•ฑ ํ™ˆ์˜ ์ž์‚ฐ ๋ฉ”๋‰ด์—์„œ ์ง€๊ฐ‘์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ๊ณ  ์ด๋”๋ฆฌ์›€(ETH)๊ณผ ํ…Œ๋”(USDT), USDC ์„ธ ๊ฐ€์ง€๋ฅผ ์ฃผ๊ณ ๋ฐ›์„ ์ˆ˜ ์žˆ๋‹ค. ์‚ผ์ฒœ๋งŒ ๋ช…์ด ์“ฐ๋Š” ๊ธˆ์œต ์•ฑ์— ์ฝ”์ธ ์ง€๊ฐ‘์ด ๋ถ™๋Š”๋‹ค๋Š” ์ ์—์„œ ์ฒ˜์Œ ์ง€๊ฐ‘์„ ์จ ๋ณด๋Š” ์‚ฌ๋žŒ์ด ๊ฝค ๋งŽ์ด ๋“ค์–ด์˜ฌ ๊ฒƒ์œผ๋กœ ๋ณธ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์ผ๋ฐ˜ ๊ณ„์ขŒ์ด์ฒด์ฒ˜๋Ÿผ ์ƒ๊ธด ํ™”๋ฉด ๋’ค์—๋Š” ๋ธ”๋ก์ฒด์ธ ์ง€๊ฐ‘์˜ ๊ทœ์น™์ด ๊ทธ๋Œ€๋กœ ๊น”๋ ค ์žˆ์–ด์„œ ์“ฐ๊ธฐ ์ „์— ์งš์–ด ๋‘˜ ๋ถ€๋ถ„์„ ๊ณต์‹ ์•ˆ๋‚ด์™€ ๋ณด๋„ ๋‚ด์šฉ์„ ๊ธฐ์ค€์œผ๋กœ ์ •๋ฆฌํ•ด ๋‘”๋‹ค. ์ด๋”๋ฆฌ์›€ ๋„คํŠธ์›Œํฌ ํ•˜๋‚˜๋งŒ ๋œ๋‹ค ์ถœ์‹œ ์‹œ์ ์— ์ง€์›ํ•˜๋Š” ๋„คํŠธ์›Œํฌ๋Š” ์ด๋”๋ฆฌ์›€ ํ•˜๋‚˜๋‹ค. ํ…Œ๋”์™€ USDC๋Š” ํŠธ๋ก ์ด๋‚˜ ์†”๋ผ๋‚˜ ๊ฐ™์€ ์—ฌ๋Ÿฌ ์ฒด์ธ์—์„œ ๋ฐœํ–‰๋˜๋Š”๋ฐ ํ† ์Šค ์ฝ”์ธ์ง€๊ฐ‘์œผ๋กœ ๋ฐ›์„ ์ˆ˜ ์žˆ๋Š” ๊ฑด ์ด๋”๋ฆฌ์›€ ๋„คํŠธ์›Œํฌ๋กœ ๋ณด๋‚ธ ๊ฒƒ๋ฟ์ด๋‹ค. ๊ฑฐ๋ž˜์†Œ์—์„œ ํ…Œ๋”๋ฅผ ์ถœ๊ธˆํ•  ๋•Œ ์ถœ๊ธˆ ๋„คํŠธ์›Œํฌ๋ฅผ ๊ณ ๋ฅด๋Š” ํ™”๋ฉด์ด ๋‚˜์˜ค๋ฉด ์ด๋”๋ฆฌ์›€(ERC-20)์„ ๊ณจ๋ผ์•ผ ํ•œ๋‹ค๋Š” ์–˜๊ธฐ๋‹ค. ์ˆ˜์ˆ˜๋ฃŒ๊ฐ€ ์‹ธ๋‹ค๊ณ  ๋‹ค๋ฅธ ๋„คํŠธ์›Œํฌ๋ฅผ ๊ณ ๋ฅด๋ฉด ํ† ์Šค ์ฝ”์ธ์ง€๊ฐ‘์—์„œ๋Š” ๋ฐ›์„ ์ˆ˜ ์—†๋‹ค. ์ง€์› ์ž์‚ฐ๊ณผ ๋„คํŠธ์›Œํฌ๋Š” ์„œ๋น„์Šค๊ฐ€ ์•ˆ์ •๋˜๋ฉด ๋‹จ๊ณ„์ ์œผ๋กœ ๋Š˜๋ฆฐ๋‹ค๊ณ  ํ–ˆ์œผ๋‹ˆ ๊ทธ์ „๊นŒ์ง€๋Š” ์ด ์ œ์•ฝ์„ ์ „์ œ๋กœ ์จ์•ผ ํ•œ๋‹ค. ํ† ์Šค ์ˆ˜์ˆ˜๋ฃŒ๋Š” ์—†์ง€๋งŒ ๊ฐ€์Šค๋น„๋Š” ETH๋กœ ๋‚ธ๋‹ค ํ† ์Šค๊ฐ€ ๋”ฐ๋กœ ๋ฐ›๋Š” ์ˆ˜์ˆ˜๋ฃŒ๋Š” ์—†๋‹ค. ๋Œ€์‹  ์™ธ๋ถ€๋กœ ๋ณด๋‚ผ ๋•Œ๋งˆ๋‹ค ์ด๋”๋ฆฌ์›€ ๋„คํŠธ์›Œํฌ ๊ฐ€์Šค๋น„๊ฐ€ ๋“ค๊ณ , ํ…Œ๋”๋‚˜ USDC๋ฅผ ๋ณด๋‚ผ ๋•Œ๋„ ๊ฐ€์Šค๋น„๋Š” ์ด๋”๋ฆฌ์›€์œผ๋กœ ๋‚ธ๋‹ค. ๊ทธ๋ž˜์„œ ํ…Œ๋”๋งŒ ๋ฐ›์•„ ๋‘” ์ง€๊ฐ‘์€ ๊ทธ ํ…Œ๋”๋ฅผ ๋‹ค์‹œ ๋ฐ–์œผ๋กœ ๋ณด๋‚ผ ์ˆ˜๊ฐ€ ์—†๋‹ค. ์ง€๊ฐ‘ ์•ˆ์— ๊ฐ€์Šค๋น„๋กœ ์“ธ ETH๊ฐ€ ์กฐ๊ธˆ์ด๋ผ๋„ ์žˆ์–ด์•ผ ํ•œ๋‹ค. ๊ฐ€์Šค๋น„๋Š” ๋ณด๋‚ด๋Š” ๊ธˆ์•ก๊ณผ ์ƒ๊ด€์—†์ด ๋ถ™์œผ๋‹ˆ ์†Œ์•ก์„ ์ž์ฃผ ์˜ฎ๊ธธ์ˆ˜๋ก ๋ถ€๋‹ด์ด ์ปค์ง„๋‹ค. ์—ฐ๋ฝ์ฒ˜ ์†ก๊ธˆ์€ ํ† ์Šค ์ด์šฉ์ž๋ผ๋ฆฌ๋งŒ ํ† ์Šค ์ด์šฉ์ž๋ผ๋ฆฌ๋Š” ๊ธด ์ง€๊ฐ‘ ์ฃผ์†Œ ๋Œ€์‹  ์—ฐ๋ฝ์ฒ˜๋กœ ๋ณด๋‚ผ ์ˆ˜ ์žˆ๋‹ค. ์ƒ๋Œ€๊ฐ€ ํ† ์Šค ์ด์šฉ์ž๊ฐ€ ์•„๋‹ˆ๋ฉด ์ง€๊ฐ‘ ์ฃผ์†Œ๋ฅผ ์ง์ ‘ ๋„ฃ์–ด์•ผ ํ•œ๋‹ค. ํ™”๋ฉด์€ ๊ณ„์ขŒ์ด์ฒด์™€ ๋‹ฎ์•˜์–ด๋„ ์™ธ๋ถ€๋กœ ๋‚˜๊ฐ„ ์†ก๊ธˆ์€ ๋ธ”๋ก์ฒด์ธ ๊ฑฐ๋ž˜๋กœ ์ฒ˜๋ฆฌ๋œ๋‹ค. ๊ฑฐ๋ž˜์†Œ์—์„œ ๋„ฃ์„ ๋•Œ๋Š” ์‚ฌ์ „ ๋“ฑ๋ก์ด ๋จผ์ €๋‹ค ๊ตญ๋‚ด ๊ฑฐ๋ž˜์†Œ์—์„œ ํ† ์Šค ์ฝ”์ธ์ง€๊ฐ‘์œผ๋กœ ๋ณด๋‚ด๋Š” ๊ธธ์€ ๋น—์ธ๊ณผ ์ฝ”์ธ์› ์•ˆ๋‚ด๊ฐ€ ๋จผ์ € ๋‚˜์™”๋‹ค. ๋น—์ธ์€ ํ† ์Šค ์ฝ”์ธ์ง€๊ฐ‘ ์ฃผ์†Œ๋ฅผ ๋ฏธ๋ฆฌ ๋“ฑ๋กํ•˜๊ณ  ์‹ฌ์‚ฌ๋ฅผ ํ†ต๊ณผํ•ด์•ผ ํ•œ๋‹ค. ๊ทธ๋ž˜๋„ ๋ฐฑ๋งŒ ์› ๋ฏธ๋งŒ๊นŒ์ง€๋งŒ ๋ณด๋‚ผ ์ˆ˜ ์žˆ๊ณ  ๊ทธ ์ด์ƒ์€ ์•„์ง ์ง€์›ํ•˜์ง€ ์•Š๋Š”๋‹ค. ์ฝ”์ธ์›๋„ ์ฃผ์†Œ๋ฅผ ์‚ฌ์ „ ๋“ฑ๋กํ•˜๊ณ  ์Šน์ธ์„ ๋ฐ›์•„์•ผ ๋ณด๋‚ผ ์ˆ˜ ์žˆ๋‹ค. ๋น„์ˆ˜ํƒ์ธ๋ฐ ํ‚ค๋Š” ํ† ์Šค๊ฐ€ ์ง€ํ‚จ๋‹ค๋Š” ๊ตฌ์กฐ ์ด ๋ถ€๋ถ„์ด ์ œ์ผ ํ—ท๊ฐˆ๋ฆฐ๋‹ค. ํ† ์Šค๋Š” ์ž์‚ฐ์„ ๋Œ€์‹  ๋ณด๊ด€ํ•˜์ง€ ์•Š๋Š” ๋น„์ˆ˜ํƒ ์ง€๊ฐ‘์ด๋ผ๊ณ  ์„ค๋ช…ํ•˜๋ฉด์„œ๋„ ์•ฑ FAQ์—์„œ ๊ฐœ์ธ ํ‚ค๋ฅผ ํ† ์Šค๊ฐ€ ์•ˆ์ „ํ•˜๊ฒŒ ๋ณดํ˜ธํ•˜๊ณ  ์žˆ๊ณ  ํœด๋Œ€์ „ํ™”๋ฅผ ์žƒ์–ด๋ฒ„๋ ค๋„ ์ƒˆ ํœด๋Œ€์ „ํ™”์— ํ† ์Šค๋ฅผ ์„ค์น˜ํ•˜๋ฉด ๊ธฐ์กด ์ง€๊ฐ‘์„ ๋‹ค์‹œ ์“ธ ์ˆ˜ ์žˆ๋‹ค๊ณ  ์•ˆ๋‚ดํ–ˆ๋‹ค. ๋ฉ”ํƒ€๋งˆ์Šคํฌ์ฒ˜๋Ÿผ ๋ณต๊ตฌ ๋ฌธ๊ตฌ๋ฅผ ์ง์ ‘ ์ ์–ด ๋‘๊ณ  ๊ทธ ๋ฌธ๊ตฌ๋กœ ์ง€๊ฐ‘์„ ๋˜์‚ด๋ฆฌ๋Š” ๋ฐฉ์‹๊ณผ๋Š” ๊ฒฐ์ด ๋‹ค๋ฅด๋‹ค. ์ง€๊ฐ‘ ๋ณต๊ตฌ๊ฐ€ ํ† ์Šค ๊ณ„์ •๊ณผ ํ† ์Šค ์•ฑ์— ๋ฌถ์—ฌ ์žˆ๋Š” ์…ˆ์ด๋ผ ํ† ์Šค ๊ณ„์ • ์ ‘๊ทผ์„ ์žƒ์œผ๋ฉด ์ง€๊ฐ‘ ์ ‘๊ทผ๋„ ๊ฐ™์ด ํ”๋“ค๋ฆฐ๋‹ค. ๋””์ง€ํ„ธ์ž์‚ฐ์ด๋ผ ์˜ˆ๊ธˆ์ž๋ณดํ˜ธ๋„ ์ ์šฉ๋˜์ง€ ์•Š๋Š”๋‹ค. ํ† ์Šค๋Š” ์•ž์œผ๋กœ ๊ฐ€์ƒ์ž์‚ฐ์‚ฌ์—…์ž ๋ผ์ด์„ ์Šค ์ทจ๋“๊ณผ ํ•จ๊ป˜ ๊ทœ์ œ ํ™˜๊ฒฝ์— ๋”ฐ๋ผ ์ˆ˜ํƒํ˜• ์ง€๊ฐ‘ ์ „ํ™˜๋„ ๊ฒ€ํ† ํ•˜๊ณ  ์žˆ๋‹ค๊ณ  ๋ฐํ˜”๋‹ค. ์ง€๊ธˆ์€ ์†ก๊ธˆ๊ณผ ์ˆ˜์‹ , ์ž”์•ก ํ™•์ธ ๊ฐ™์€ ๊ธฐ๋ณธ ๊ธฐ๋Šฅ๋งŒ ์žˆ๋Š” ์ฒซ ๋ฒ„์ „์ด๋‹ˆ ํฐ ๊ธˆ์•ก์„ ๋ฌตํ˜€ ๋‘๋Š” ์ง€๊ฐ‘๋ณด๋‹ค ์Šคํ…Œ์ด๋ธ”์ฝ”์ธ์„ ์ฃผ๊ณ ๋ฐ›๋Š” ์ฐฝ๊ตฌ ์ •๋„๋กœ ๋ณด๋Š” ํŽธ์ด ๊ตฌ์กฐ์— ๋งž๋‹ค๊ณ  ๋ณธ๋‹ค.
2
5
189
๐Ÿ“š Rialo Research Series 64: Agent Shutdown Authority - Who Can Stop an Autonomous Agent? (Kor : ์—์ด์ „ํŠธ ์ค‘๋‹จ ๊ถŒํ•œ - ์ž์œจ์ ์œผ๋กœ ์›€์ง์ด๋Š” Agent๋ฅผ ๋ˆ„๊ฐ€ ๋ฉˆ์ถœ ์ˆ˜ ์žˆ์„๊นŒ?) 1/ Intro: Accountability Is Not Enough (EN) In the previous chapter, we explored how responsibility could be traced across multiple agents involved in a failed financial decision. But identifying responsibility after an incident doesn't prevent further damage while the incident is unfolding. Imagine an autonomous trading agent repeatedly executing transactions based on corrupted market data. Investigators may eventually determine what went wrong. Meanwhile, the agent continues trading. Accountability explains a failure. Shutdown authority determines whether the failure can be contained. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ์—ฌ๋Ÿฌ Agent๊ฐ€ ๊ด€์—ฌํ•œ ์ž˜๋ชป๋œ ๊ธˆ์œต ์˜์‚ฌ๊ฒฐ์ •์—์„œ ์ฑ…์ž„์„ ์–ด๋–ป๊ฒŒ ์ถ”์ ํ•  ์ˆ˜ ์žˆ์„์ง€ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์‚ฌ๊ณ  ์ดํ›„ ์ฑ…์ž„์„ ํ™•์ธํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ๋Š” ์ง„ํ–‰ ์ค‘์ธ ํ”ผํ•ด๋ฅผ ๋ง‰์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. Autonomous Trading Agent๊ฐ€ ์˜ค์—ผ๋œ Market Data๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ณ„์† ๊ฑฐ๋ž˜ํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋‚˜์ค‘์— ์‚ฌ๊ณ  ์›์ธ์„ ๋ฐํ˜€๋‚ด๋”๋ผ๋„ ๊ทธ๋™์•ˆ Agent๋Š” ๊ฑฐ๋ž˜๋ฅผ ๊ณ„์† ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Accountability๊ฐ€ ์‹คํŒจ๋ฅผ ์„ค๋ช…ํ•˜๋Š” ๊ณผ์ •์ด๋ผ๋ฉด Shutdown Authority๋Š” ํ”ผํ•ด ํ™•์‚ฐ์„ ๋ง‰๋Š” ํ†ต์ œ ์ˆ˜๋‹จ์ž…๋‹ˆ๋‹ค. 2/ The Problem With Autonomous Execution (EN) Traditional financial systems generally have operators who can suspend services or restrict transactions. Autonomous agents complicate this arrangement. An agent may operate continuously, interact with multiple protocols, and trigger transactions without requesting approval for every action. Disabling its user interface may not stop already authorized workflows or pending transactions. Effective shutdown mechanisms must therefore operate at the level where execution authority is actually enforced. (KR) ์ „ํ†ต์ ์ธ ๊ธˆ์œต ์‹œ์Šคํ…œ์—์„œ๋Š” ์šด์˜์ž๊ฐ€ ์„œ๋น„์Šค๋ฅผ ์ค‘๋‹จํ•˜๊ฑฐ๋‚˜ ๊ฑฐ๋ž˜๋ฅผ ์ œํ•œํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Autonomous Agent๋Š” ์ƒํ™ฉ์ด ์กฐ๊ธˆ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. Agent๋Š” ์ง€์†์ ์œผ๋กœ ์ž‘๋™ํ•˜๋ฉด์„œ ์—ฌ๋Ÿฌ Protocol๊ณผ ์ƒํ˜ธ์ž‘์šฉํ•˜๊ณ , ๋งค๋ฒˆ ์Šน์ธ์„ ๋ฐ›์ง€ ์•Š๊ณ  Transaction์„ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹จ์ˆœํžˆ ์‚ฌ์šฉ์ž ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ๋น„ํ™œ์„ฑํ™”ํ•˜๋”๋ผ๋„ ์ด๋ฏธ ์Šน์ธ๋œ Workflow๋‚˜ Pending Transaction๊นŒ์ง€ ์ค‘๋‹จ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์‹ค์ œ ์‹คํ–‰ ๊ถŒํ•œ์ด ์ ์šฉ๋˜๋Š” ์ง€์ ์—์„œ ํ†ต์ œํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 3/ Three Levels of Intervention (EN) Not every incident requires permanently disabling an agent. A system could distinguish between three intervention levels. The first restricts specific actions, such as borrowing or liquidation. The second temporarily suspends the agent's execution permissions while preserving its records and state. The third revokes critical authorization when an agent is compromised or cannot be safely recovered. These controls should have clearly defined triggers and authorization requirements. (KR) ๋ชจ๋“  ์‚ฌ๊ณ ์—์„œ Agent๋ฅผ ์˜๊ตฌ์ ์œผ๋กœ ์ค‘๋‹จํ•  ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. ๊ฐœ์ž… ์ˆ˜์ค€์„ ์„ธ ๋‹จ๊ณ„๋กœ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋Š” Borrowing์ด๋‚˜ Liquidation ๊ฐ™์€ ํŠน์ • ํ–‰๋™๋งŒ ์ œํ•œํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๋‘ ๋ฒˆ์งธ๋Š” Agent์˜ ๊ธฐ๋ก๊ณผ ์ƒํƒœ๋ฅผ ์œ ์ง€ํ•˜๋ฉด์„œ ์‹คํ–‰ ๊ถŒํ•œ์„ ์ผ์‹œ์ ์œผ๋กœ ์ •์ง€ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์„ธ ๋ฒˆ์งธ๋Š” Agent๊ฐ€ ํƒˆ์ทจ๋๊ฑฐ๋‚˜ ์•ˆ์ „ํ•˜๊ฒŒ ๋ณต๊ตฌํ•˜๊ธฐ ์–ด๋ ค์šด ๊ฒฝ์šฐ ํ•ต์‹ฌ Authorization์„ ์ฒ ํšŒํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ๊ฐ ๋‹จ๊ณ„๋งˆ๋‹ค ๋ฐœ๋™ ์กฐ๊ฑด๊ณผ ์Šน์ธ ๊ถŒํ•œ์ด ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜๋ผ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 4/ Who Holds the Emergency Authority? (EN) Giving one administrator unrestricted shutdown authority creates a centralization risk. But requiring a full governance vote during an active exploit could make intervention too slow. A possible compromise is separating emergency suspension from permanent policy changes. For example, a designated security committee could temporarily pause high-risk operations, while restoring permissions requires additional review. The challenge is balancing response speed with protection against authority abuse. (KR) ํ•œ ๋ช…์˜ ๊ด€๋ฆฌ์ž์—๊ฒŒ ๋ฌด์ œํ•œ์ ์ธ Shutdown Authority๋ฅผ ๋ถ€์—ฌํ•˜๋ฉด ์ค‘์•™ํ™” ์œ„ํ—˜์ด ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ์ง„ํ–‰ ์ค‘์ธ ๊ณต๊ฒฉ์„ ๋ง‰๊ธฐ ์œ„ํ•ด ๋ชจ๋“  ๊ฒฐ์ •์„ Governance Vote๋กœ ์ฒ˜๋ฆฌํ•œ๋‹ค๋ฉด ๋Œ€์‘์ด ์ง€๋‚˜์น˜๊ฒŒ ๋Šฆ์–ด์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•œ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์€ ๊ธด๊ธ‰ ์ค‘๋‹จ ๊ถŒํ•œ๊ณผ ์˜๊ตฌ์ ์ธ ์ •์ฑ… ๋ณ€๊ฒฝ ๊ถŒํ•œ์„ ๋ถ„๋ฆฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์ง€์ •๋œ Security Committee๊ฐ€ ๊ณ ์œ„ํ—˜ ์ž‘์—…์„ ์ผ์‹œ ์ค‘๋‹จํ•˜๋˜, ๊ถŒํ•œ ๋ณต๊ตฌ์—๋Š” ๋ณ„๋„์˜ ๊ฒ€ํ†  ์ ˆ์ฐจ๋ฅผ ์š”๊ตฌํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•ต์‹ฌ์€ ๋น ๋ฅธ ์‚ฌ๊ณ  ๋Œ€์‘๊ณผ ๊ถŒํ•œ ๋‚จ์šฉ ๋ฐฉ์ง€ ์‚ฌ์ด์˜ ๊ท ํ˜•์ž…๋‹ˆ๋‹ค. 5/ What Happens to Pending Transactions? (EN) Stopping an agent does not necessarily cancel everything it has already initiated. Some transactions may be waiting for inclusion, while others may already be finalized. External protocols may also retain permissions granted before the shutdown. A reliable intervention design must distinguish between preventing new instructions, revoking existing permissions, and handling transactions that can no longer be stopped. A shutdown button is only effective when its actual scope is understood. (KR) Agent๋ฅผ ์ค‘๋‹จํ•œ๋‹ค๊ณ  ํ•ด์„œ ์ด๋ฏธ ์‹œ์ž‘๋œ ๋ชจ๋“  ์ž‘์—…์ด ์ทจ์†Œ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ผ๋ถ€ Transaction์€ ์•„์ง ์ฒ˜๋ฆฌ ๋Œ€๊ธฐ ์ค‘์ผ ์ˆ˜ ์žˆ๊ณ , ๋‹ค๋ฅธ Transaction์€ ์ด๋ฏธ Finalized๋์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์™ธ๋ถ€ Protocol์— ๋ถ€์—ฌํ•œ ๊ถŒํ•œ ์—ญ์‹œ ์ค‘๋‹จ ์ดํ›„์—๋„ ๋‚จ์•„ ์žˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ƒˆ๋กœ์šด ๋ช…๋ น ์ฐจ๋‹จ, ๊ธฐ์กด ๊ถŒํ•œ ์ฒ ํšŒ, ์ค‘๋‹จํ•  ์ˆ˜ ์—†๋Š” ๊ฑฐ๋ž˜์˜ ์‚ฌํ›„ ์ฒ˜๋ฆฌ๊นŒ์ง€ ๊ตฌ๋ถ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Shutdown ๊ธฐ๋Šฅ์˜ ์‹คํšจ์„ฑ์€ ๋ฒ„ํŠผ์˜ ์กด์žฌ๊ฐ€ ์•„๋‹ˆ๋ผ ์‹ค์ œ๋กœ ์–ด๋””๊นŒ์ง€ ํ†ต์ œํ•  ์ˆ˜ ์žˆ๋А๋ƒ์— ๋‹ฌ๋ ค ์žˆ์Šต๋‹ˆ๋‹ค. 6/ Rialo and Reactive Safety Controls (EN) Rialo's Reactive Transactions provide an interesting foundation for designing condition-based safety responses. An application could monitor predefined risk conditions and trigger restrictions when those conditions are satisfied. For example, repeated oracle inconsistencies might activate a temporary limitation on liquidation operations. However, reactive execution does not automatically grant universal authority to stop agents or reverse transactions. The application must still implement permission boundaries, emergency roles, and recovery procedures. (KR) Rialo์˜ Reactive Transactions๋Š” ์กฐ๊ฑด ๊ธฐ๋ฐ˜ Safety Control์„ ์„ค๊ณ„ํ•˜๋Š” ๋ฐ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ๋ฐ˜์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. Application์€ ์‚ฌ์ „์— ์ •์˜ํ•œ Risk Condition์„ ๊ฐ์‹œํ•˜๊ณ , ํ•ด๋‹น ์กฐ๊ฑด์ด ์ถฉ์กฑ๋˜๋ฉด ์ œํ•œ ์กฐ์น˜๋ฅผ ์‹คํ–‰ํ•˜๋„๋ก ๊ตฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Oracle ๋ฐ์ดํ„ฐ ๋ถˆ์ผ์น˜๊ฐ€ ๋ฐ˜๋ณต๋˜๋ฉด Liquidation ๊ธฐ๋Šฅ์„ ์ผ์‹œ์ ์œผ๋กœ ์ œํ•œํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ Reactive Execution์ด ๋ชจ๋“  Agent๋ฅผ ์ค‘๋‹จํ•˜๊ฑฐ๋‚˜ Transaction์„ ๋˜๋Œ๋ฆด ์ˆ˜ ์žˆ๋Š” ๋ณดํŽธ์ ์ธ ๊ถŒํ•œ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์‹ค์ œ Permission Boundary, Emergency Role, Recovery Procedure๋Š” Application์—์„œ ๊ตฌํ˜„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a demonstration involving an autonomous lending agent operating under normal market conditions. During execution, the oracle begins returning inconsistent prices. The system detects repeated anomalies and automatically restricts new liquidation instructions while preserving the agent's decision history. An authorized operator then reviews the incident and decides whether execution can safely resume. The demonstration should also show which pending transactions remain unaffected by the shutdown. That would reveal the practical limits of automated intervention rather than presenting shutdown as an absolute guarantee. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์ •์ƒ์ ์ธ ์‹œ์žฅ ํ™˜๊ฒฝ์—์„œ ์ž‘๋™ํ•˜๋Š” Autonomous Lending Agent Demo์ž…๋‹ˆ๋‹ค. ์‹คํ–‰ ๋„์ค‘ Oracle์ด ์„œ๋กœ ์ผ์น˜ํ•˜์ง€ ์•Š๋Š” ๊ฐ€๊ฒฉ์„ ๋ฐ˜ํ™˜ํ•˜๊ธฐ ์‹œ์ž‘ํ•ฉ๋‹ˆ๋‹ค. ์‹œ์Šคํ…œ์€ ๋ฐ˜๋ณต๋˜๋Š” ์ด์ƒ ์ง•ํ›„๋ฅผ ๊ฐ์ง€ํ•˜๊ณ , Agent์˜ Decision History๋ฅผ ๋ณด์กดํ•˜๋ฉด์„œ ์ƒˆ๋กœ์šด Liquidation ๋ช…๋ น์„ ์ž๋™์œผ๋กœ ์ œํ•œํ•ฉ๋‹ˆ๋‹ค. ์ดํ›„ ๊ถŒํ•œ์„ ๊ฐ€์ง„ ์šด์˜์ž๊ฐ€ ์‚ฌ๊ณ ๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ์‹คํ–‰ ์žฌ๊ฐœ ์—ฌ๋ถ€๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค. ๋™์‹œ์— Shutdown ์ดํ›„์—๋„ ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š๋Š” Pending Transaction์ด ๋ฌด์—‡์ธ์ง€ ๋ณด์—ฌ์ฃผ๋ฉด ์ข‹๊ฒ ์Šต๋‹ˆ๋‹ค. ์ž๋™ ์ค‘๋‹จ์ด ๋ชจ๋“  ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•œ๋‹ค๋Š” ์ฃผ์žฅ๋ณด๋‹ค ์‹ค์ œ ํ†ต์ œ ๋ฒ”์œ„์™€ ํ•œ๊ณ„๋ฅผ ๊ฒ€์ฆํ•˜๋Š” ๊ณผ์ •์ด ๋” ์ค‘์š”ํ•˜๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. (EN) Autonomous finance needs more than agents capable of making independent decisions. It also needs mechanisms that can intervene when those decisions become dangerous. The ability to suspend execution, revoke permissions, and recover safely should be designed before an incident occurs, not improvised afterward. The real measure of autonomy may not be how long an agent can operate without humans, but how safely humans can regain control when necessary. (KR) Autonomous Finance์—๋Š” ๋…๋ฆฝ์ ์œผ๋กœ ํŒ๋‹จํ•˜๊ณ  ์‹คํ–‰ํ•˜๋Š” Agent๋งŒ ํ•„์š”ํ•œ ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ ํŒ๋‹จ์ด ์œ„ํ—˜ํ•ด์กŒ์„ ๋•Œ ๊ฐœ์ž…ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ๋„ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ์‹คํ–‰ ์ค‘๋‹จ, ๊ถŒํ•œ ์ฒ ํšŒ, ์•ˆ์ „ํ•œ ๋ณต๊ตฌ ์ ˆ์ฐจ๋Š” ์‚ฌ๊ณ ๊ฐ€ ๋ฐœ์ƒํ•œ ๋’ค ๊ณ ๋ฏผํ•  ๋ฌธ์ œ๊ฐ€ ์•„๋‹ˆ๋ผ ์‚ฌ์ „์— ์„ค๊ณ„ํ•ด์•ผ ํ•  ์š”์†Œ์ž…๋‹ˆ๋‹ค. Agent๊ฐ€ ์–ผ๋งˆ๋‚˜ ์˜ค๋žซ๋™์•ˆ ์‚ฌ๋žŒ ์—†์ด ์ž‘๋™ํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋„ ์ค‘์š”ํ•˜๊ฒ ์ง€๋งŒ, ์œ„ํ—˜ํ•œ ์ƒํ™ฉ์—์„œ ์‚ฌ๋žŒ์ด ์–ผ๋งˆ๋‚˜ ์•ˆ์ „ํ•˜๊ฒŒ ํ†ต์ œ๊ถŒ์„ ๋˜์ฐพ์„ ์ˆ˜ ์žˆ๋Š”์ง€๋„ ์ค‘์š”ํ•˜์ง€ ์•Š์„๊นŒ์š”? ์ง„์ •ํ•œ ์ž์œจ์„ฑ์€ ํ†ต์ œํ•  ์ˆ˜ ์—†๋Š” ์ž์œ ๊ฐ€ ์•„๋‹ˆ๋ผ ํ•„์š”ํ•  ๋•Œ ์•ˆ์ „ํ•˜๊ฒŒ ๊ฐœ์ž…ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ ์œ„์—์„œ ์™„์„ฑ๋œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
๐Ÿ“š Rialo Research Series 63: Multi-Agent Accountability - Who Is Responsible When Autonomous Finance Fails? (Kor : ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์ฑ…์ž„ ๊ตฌ์กฐ - ์ž์œจ ๊ธˆ์œต ์‹œ์Šคํ…œ์—์„œ ์—ฌ๋Ÿฌ Agent๊ฐ€ ์ž˜๋ชป๋œ ๊ฒฐ์ •์„ ๋‚ด๋ ธ๋‹ค๋ฉด ๋ˆ„๊ตฌ์˜ ์ฑ…์ž„์ผ๊นŒ?) 1/ Intro: One Failure, Multiple Agents (EN) In the previous chapter, we explored how financial losses caused by incorrect automated decisions might be compensated. But what happens when several agents contribute to the same failure? Imagine three agents managing a lending position. One collects oracle prices, another evaluates liquidation risk, and the third executes transactions. The position is incorrectly liquidated. Each agent performed its assigned task, yet the overall outcome was wrong. When responsibility is distributed across agents, identifying the source of failure becomes a separate challenge. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ์ž˜๋ชป๋œ ์ž๋™ํ™” ํŒ๋‹จ์œผ๋กœ ๋ฐœ์ƒํ•œ ๊ธˆ์œต ์†์‹ค์„ ์–ด๋–ป๊ฒŒ ๋ณด์ƒํ•  ์ˆ˜ ์žˆ์„์ง€ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ํ•˜๋‚˜์˜ ์ž˜๋ชป๋œ ๊ฒฐ์ •์— ์—ฌ๋Ÿฌ Agent๊ฐ€ ๊ด€์—ฌํ–ˆ๋‹ค๋ฉด ์–ด๋–จ๊นŒ์š”? ์„ธ Agent๊ฐ€ Lending Position์„ ๊ด€๋ฆฌํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋Š” Oracle ๊ฐ€๊ฒฉ์„ ์ˆ˜์ง‘ํ•˜๊ณ , ๋‘ ๋ฒˆ์งธ๋Š” Liquidation Risk๋ฅผ ํ‰๊ฐ€ํ•˜๋ฉฐ, ์„ธ ๋ฒˆ์งธ๋Š” Transaction์„ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ ์ •์ƒ์ ์ธ ํฌ์ง€์…˜์ด ์ž˜๋ชป ์ฒญ์‚ฐ๋์Šต๋‹ˆ๋‹ค. ๊ฐ Agent๋Š” ์ž์‹ ์˜ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ–ˆ์ง€๋งŒ ์ตœ์ข… ๊ฒฐ๊ณผ๋Š” ์ž˜๋ชป๋์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ Agent๊ฐ€ ํ˜‘๋ ฅํ•˜๋Š” ๊ตฌ์กฐ์—์„œ๋Š” ์‹คํŒจ ์›์ธ์„ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ๋ถ€ํ„ฐ ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. 2/ Execution Responsibility vs Decision Responsibility (EN) The agent executing a transaction is not necessarily the agent responsible for the incorrect decision. An execution agent may faithfully follow instructions generated by a risk assessment agent. Meanwhile, the risk assessment itself may rely on inaccurate information from another source. Simply identifying the wallet that submitted the transaction cannot explain the entire decision process. We need to distinguish who supplied information, who interpreted it, and who authorized execution. (KR) Transaction์„ ์‹คํ–‰ํ•œ Agent๊ฐ€ ๋ฐ˜๋“œ์‹œ ์ž˜๋ชป๋œ ํŒ๋‹จ์˜ ์ฑ…์ž„์ž๋ผ๊ณ  ๋ณผ ์ˆ˜๋Š” ์—†์Šต๋‹ˆ๋‹ค. Execution Agent๋Š” Risk Assessment Agent๊ฐ€ ์ „๋‹ฌํ•œ ๋ช…๋ น์„ ์ •ํ™•ํ•˜๊ฒŒ ์ˆ˜ํ–‰ํ–ˆ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ Risk Assessment ์ž์ฒด๊ฐ€ ๋‹ค๋ฅธ Agent์˜ ๋ถ€์ •ํ™•ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ด๋ฃจ์–ด์กŒ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Transaction์„ ์ œ์ถœํ•œ Wallet๋งŒ ํ™•์ธํ•ด์„œ๋Š” ์ „์ฒด ์˜์‚ฌ๊ฒฐ์ • ๊ณผ์ •์„ ์„ค๋ช…ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ •๋ณด ์ œ๊ณต, ํŒ๋‹จ, ์‹คํ–‰ ์Šน์ธ์ด๋ผ๋Š” ์—ญํ• ์„ ๊ฐ๊ฐ ๊ตฌ๋ถ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 3/ Reconstructing the Decision Chain (EN) Suppose Agent A provides an incorrect collateral price. Agent B uses that price to calculate a liquidation threshold, and Agent C executes the liquidation. To investigate the incident, we would need to reconstruct the decision chain. What information did Agent B receive? Was the data marked as uncertain? Did Agent C verify the authorization conditions? Transaction logs alone may not capture every off-chain decision. Applications need reliable records connecting inputs, decisions, permissions, and execution results. (KR) Agent A๊ฐ€ ์ž˜๋ชป๋œ ๋‹ด๋ณด ๊ฐ€๊ฒฉ์„ ์ œ๊ณตํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. Agent B๋Š” ํ•ด๋‹น ๊ฐ€๊ฒฉ์œผ๋กœ Liquidation Threshold๋ฅผ ๊ณ„์‚ฐํ•˜๊ณ , Agent C๋Š” ์ฒญ์‚ฐ์„ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ๊ณ  ์›์ธ์„ ํ™•์ธํ•˜๋ ค๋ฉด ์ „์ฒด Decision Chain์„ ์žฌ๊ตฌ์„ฑํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Agent B๋Š” ์–ด๋–ค ์ •๋ณด๋ฅผ ๋ฐ›์•˜๋Š”์ง€, ๋ฐ์ดํ„ฐ์˜ ๋ถˆํ™•์‹ค์„ฑ์„ ์ธ์ง€ํ–ˆ๋Š”์ง€, Agent C๋Š” ์‹คํ–‰ ๊ถŒํ•œ์„ ๊ฒ€์ฆํ–ˆ๋Š”์ง€ ํ™•์ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. On-chain Transaction Log๋งŒ์œผ๋กœ๋Š” ๋ชจ๋“  Off-chain ํŒ๋‹จ์„ ํŒŒ์•…ํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Input, Decision, Permission, Execution Result๋ฅผ ์—ฐ๊ฒฐํ•˜๋Š” ๊ธฐ๋ก์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 4/ When Multiple Agents Share Responsibility (EN) Not every failure has a single cause. Agent A might provide outdated data, while Agent B ignores a freshness warning. Agent C could then execute the transaction without checking an additional safety condition. In this situation, assigning all responsibility to one agent would overlook the other failures. A useful accountability framework should distinguish the original error from failures that allowed the error to propagate. However, technical fault attribution does not automatically determine legal or financial liability. (KR) ๋ชจ๋“  ์‚ฌ๊ณ ์˜ ์›์ธ์ด ํ•˜๋‚˜์ธ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. Agent A๊ฐ€ ์˜ค๋ž˜๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ–ˆ๊ณ , Agent B๋Š” ๋ฐ์ดํ„ฐ ๊ฐฑ์‹  ๊ฒฝ๊ณ ๋ฅผ ๋ฌด์‹œํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. Agent C ์—ญ์‹œ ์ถ”๊ฐ€์ ์ธ Safety Condition์„ ํ™•์ธํ•˜์ง€ ์•Š๊ณ  Transaction์„ ์‹คํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ ํ•˜๋‚˜์˜ Agent์— ๋ชจ๋“  ์ฑ…์ž„์„ ๋ถ€์—ฌํ•˜๋ฉด ๋‹ค๋ฅธ ๋ฌธ์ œ๋“ค์„ ๋†“์น˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ตœ์ดˆ ์˜ค๋ฅ˜์™€ ์˜ค๋ฅ˜๊ฐ€ ํ™•์‚ฐ๋˜๋Š” ๊ฒƒ์„ ๋ง‰์ง€ ๋ชปํ•œ ์‹คํŒจ๋ฅผ ๊ตฌ๋ถ„ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ ๊ธฐ์ˆ ์ ์ธ ์˜ค๋ฅ˜ ๊ธฐ์—ฌ๋„๋ฅผ ํ™•์ธํ•˜๋Š” ๊ฒƒ๊ณผ ๋ฒ•์ ยท๊ธˆ์ „์  ์ฑ…์ž„์„ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. 5/ Preventing the Blame Game (EN) Without verifiable records, agents could produce conflicting explanations after an incident. One agent might claim that another supplied incorrect data. Another might argue that its instructions were misunderstood. This creates a difficult problem when autonomous systems operate without continuous human supervision. Signed messages, versioned inputs, authorization records, and reproducible execution traces could help establish what actually happened. The objective is not to make every agent liable, but to prevent responsibility from disappearing between them. (KR) ๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ๊ธฐ๋ก์ด ์—†๋‹ค๋ฉด ์‚ฌ๊ณ  ๋ฐœ์ƒ ํ›„ Agent๋“ค์ด ์„œ๋กœ ๋‹ค๋ฅธ ์„ค๋ช…์„ ๋‚ด๋†“์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•œ Agent๋Š” ์ž˜๋ชป๋œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ›์•˜๋‹ค๊ณ  ์ฃผ์žฅํ•˜๊ณ , ๋‹ค๋ฅธ Agent๋Š” ์ž์‹ ์˜ ์ง€์‹œ๊ฐ€ ์ž˜๋ชป ํ•ด์„๋๋‹ค๊ณ  ์ฃผ์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒ์ด ์ง€์†์ ์œผ๋กœ ๊ฐ๋…ํ•˜์ง€ ์•Š๋Š” Autonomous System์—์„œ๋Š” ์ด๋Ÿฐ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ๋ณต์žกํ•ด์ง‘๋‹ˆ๋‹ค. Signed Message, Versioned Input, Authorization Record, Execution Trace ๋“ฑ์„ ํ™œ์šฉํ•˜๋ฉด ์‹ค์ œ ๊ณผ์ •์„ ๊ฒ€์ฆํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ํ•ต์‹ฌ์€ ๋ชจ๋“  Agent์—๊ฒŒ ์ฑ…์ž„์„ ๋ถ€๊ณผํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ Agent ์‚ฌ์ด์—์„œ ์ฑ…์ž„์˜ ๊ทผ๊ฑฐ๊ฐ€ ์‚ฌ๋ผ์ง€์ง€ ์•Š๋„๋ก ๋งŒ๋“œ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. 6/ Rialo and Verifiable Agent Coordination (EN) Rialo's execution environment provides an interesting foundation for exploring these problems. Reactive Transactions could support predefined responses when suspicious conditions are detected. REX and confidential execution could also be relevant to workflows involving sensitive agent information. However, these capabilities do not automatically establish a complete accountability framework. Applications would still need to define how agent identities, decision records, permissions, and dispute procedures are managed. Verifiable execution is an important building block, but accountability requires more than execution alone. (KR) Rialo์˜ Execution Environment๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํƒ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ํฅ๋ฏธ๋กœ์šด ๊ธฐ๋ฐ˜์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Reactive Transactions๋ฅผ ํ™œ์šฉํ•˜๋ฉด ์ด์ƒ ์กฐ๊ฑด์ด ๊ฐ์ง€๋์„ ๋•Œ ์‚ฌ์ „์— ์ •์˜๋œ ๋Œ€์‘ ์ ˆ์ฐจ๋ฅผ ์‹คํ–‰ํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. REX์™€ Confidential Execution ์—ญ์‹œ ๋ฏผ๊ฐํ•œ Agent ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” Workflow์—์„œ ํ™œ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด๋Ÿฌํ•œ ๊ธฐ๋Šฅ๋งŒ์œผ๋กœ ์™„์ „ํ•œ Accountability Framework๊ฐ€ ์ž๋™์œผ๋กœ ๋งŒ๋“ค์–ด์ง€๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. Agent Identity, Decision Record, Permission, Dispute Procedure ๋“ฑ์„ Application ์ฐจ์›์—์„œ ์ •์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ์‹คํ–‰์€ ์ค‘์š”ํ•œ ์ถœ๋ฐœ์ ์ด์ง€๋งŒ ์ฑ…์ž„ ๊ตฌ์กฐ๋ฅผ ์™„์„ฑํ•˜๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a multi-agent lending demonstration involving three independent agents. The first supplies market data, the second evaluates risk, and the third executes transactions. An intentionally outdated price causes an incorrect liquidation. The demonstration should then reconstruct the decision chain and identify whether the failure originated from bad data, ignored warnings, or insufficient execution safeguards. Ideally, the system would also demonstrate how verified evidence can initiate a controlled remediation process. This would make agent accountability a practical engineering problem rather than an abstract promise. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์„ธ ๊ฐœ์˜ ๋…๋ฆฝ์ ์ธ Agent๊ฐ€ ์ฐธ์—ฌํ•˜๋Š” Lending Demo์ž…๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ Agent๋Š” Market Data๋ฅผ ์ œ๊ณตํ•˜๊ณ , ๋‘ ๋ฒˆ์งธ๋Š” Risk๋ฅผ ํ‰๊ฐ€ํ•˜๋ฉฐ, ์„ธ ๋ฒˆ์งธ๋Š” Transaction์„ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์˜๋„์ ์œผ๋กœ ์˜ค๋ž˜๋œ ๊ฐ€๊ฒฉ ์ •๋ณด๋ฅผ ์ „๋‹ฌํ•ด ์ž˜๋ชป๋œ Liquidation์„ ๋ฐœ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ดํ›„ Decision Chain์„ ์žฌ๊ตฌ์„ฑํ•ด ๋ฐ์ดํ„ฐ ์˜ค๋ฅ˜์ธ์ง€, ๊ฒฝ๊ณ  ๋ฌด์‹œ์ธ์ง€, ์‹คํ–‰ ๋‹จ๊ณ„์˜ ์•ˆ์ „์žฅ์น˜ ๋ถ€์กฑ์ธ์ง€ ํ™•์ธํ•˜๋Š” ๊ณผ์ •์„ ๋ณด๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ๋‚˜์•„๊ฐ€ ๊ฒ€์ฆ๋œ Evidence๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Remediation Process๋ฅผ ์‹œ์ž‘ํ•˜๋Š” ๋ชจ์Šต๊นŒ์ง€ ๋ณด์—ฌ์ค€๋‹ค๋ฉด ๋”์šฑ ์˜๋ฏธ ์žˆ์„ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. Agent Accountability๊ฐ€ ์ถ”์ƒ์ ์ธ ๊ฐœ๋…์ด ์•„๋‹ˆ๋ผ ์‹ค์ œ๋กœ ๊ตฌํ˜„ํ•˜๊ณ  ๊ฒ€์ฆํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ์ˆ ์  ๊ณผ์ œ๋ผ๋Š” ์ ์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. (EN) As autonomous finance evolves, multi-agent collaboration will make financial decisions increasingly complex. A transaction may have one executor, but its outcome can depend on information and decisions produced by several independent agents. The challenge is not simply identifying who executed a transaction. It is preserving enough verifiable evidence to understand how the decision was made, where safeguards failed, and what corrective action is justified. Autonomous systems should not only be capable of acting independently. They should also make their failures accountable. (KR) Autonomous Finance๊ฐ€ ๋ฐœ์ „ํ• ์ˆ˜๋ก ์—ฌ๋Ÿฌ Agent์˜ ํ˜‘์—…์œผ๋กœ ์ด๋ฃจ์–ด์ง€๋Š” ๊ธˆ์œต ์˜์‚ฌ๊ฒฐ์ •์€ ๋”์šฑ ๋ณต์žกํ•ด์งˆ ๊ฒƒ์ž…๋‹ˆ๋‹ค. Transaction์„ ์‹คํ–‰ํ•œ ์ฃผ์ฒด๋Š” ํ•˜๋‚˜์ผ ์ˆ˜ ์žˆ์ง€๋งŒ, ๊ทธ ๊ฒฐ๊ณผ์—๋Š” ์—ฌ๋Ÿฌ ๋…๋ฆฝ์ ์ธ Agent๊ฐ€ ์ œ๊ณตํ•œ ์ •๋ณด์™€ ํŒ๋‹จ์ด ๋ฐ˜์˜๋ฉ๋‹ˆ๋‹ค. ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋ˆ„๊ฐ€ Transaction์„ ์‹คํ–‰ํ–ˆ๋Š”์ง€๋งŒ ํ™•์ธํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์–ด๋–ค ๊ณผ์ •์„ ๊ฑฐ์ณ ๊ฒฐ์ •์ด ๋‚ด๋ ค์กŒ๊ณ , ์–ด๋””์—์„œ ์•ˆ์ „์žฅ์น˜๊ฐ€ ์ž‘๋™ํ•˜์ง€ ์•Š์•˜์œผ๋ฉฐ, ์–ด๋–ค ํ›„์† ์กฐ์น˜๊ฐ€ ์ •๋‹นํ•œ์ง€ ๊ฒ€์ฆํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Autonomous System์€ ์Šค์Šค๋กœ ํ–‰๋™ํ•˜๋Š” ๋Šฅ๋ ฅ๋ฟ ์•„๋‹ˆ๋ผ ์‹คํŒจํ–ˆ์„ ๋•Œ ๊ทธ ์›์ธ๊ณผ ์ฑ…์ž„์„ ์ถ”์ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ๊นŒ์ง€ ๊ฐ–์ถฐ์•ผ ํ•˜์ง€ ์•Š์„๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
1
139
๐Ÿ“š Rialo Research Series 63: Multi-Agent Accountability - Who Is Responsible When Autonomous Finance Fails? (Kor : ๋‹ค์ค‘ ์—์ด์ „ํŠธ ์ฑ…์ž„ ๊ตฌ์กฐ - ์ž์œจ ๊ธˆ์œต ์‹œ์Šคํ…œ์—์„œ ์—ฌ๋Ÿฌ Agent๊ฐ€ ์ž˜๋ชป๋œ ๊ฒฐ์ •์„ ๋‚ด๋ ธ๋‹ค๋ฉด ๋ˆ„๊ตฌ์˜ ์ฑ…์ž„์ผ๊นŒ?) 1/ Intro: One Failure, Multiple Agents (EN) In the previous chapter, we explored how financial losses caused by incorrect automated decisions might be compensated. But what happens when several agents contribute to the same failure? Imagine three agents managing a lending position. One collects oracle prices, another evaluates liquidation risk, and the third executes transactions. The position is incorrectly liquidated. Each agent performed its assigned task, yet the overall outcome was wrong. When responsibility is distributed across agents, identifying the source of failure becomes a separate challenge. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ์ž˜๋ชป๋œ ์ž๋™ํ™” ํŒ๋‹จ์œผ๋กœ ๋ฐœ์ƒํ•œ ๊ธˆ์œต ์†์‹ค์„ ์–ด๋–ป๊ฒŒ ๋ณด์ƒํ•  ์ˆ˜ ์žˆ์„์ง€ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ํ•˜๋‚˜์˜ ์ž˜๋ชป๋œ ๊ฒฐ์ •์— ์—ฌ๋Ÿฌ Agent๊ฐ€ ๊ด€์—ฌํ–ˆ๋‹ค๋ฉด ์–ด๋–จ๊นŒ์š”? ์„ธ Agent๊ฐ€ Lending Position์„ ๊ด€๋ฆฌํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋Š” Oracle ๊ฐ€๊ฒฉ์„ ์ˆ˜์ง‘ํ•˜๊ณ , ๋‘ ๋ฒˆ์งธ๋Š” Liquidation Risk๋ฅผ ํ‰๊ฐ€ํ•˜๋ฉฐ, ์„ธ ๋ฒˆ์งธ๋Š” Transaction์„ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ ์ •์ƒ์ ์ธ ํฌ์ง€์…˜์ด ์ž˜๋ชป ์ฒญ์‚ฐ๋์Šต๋‹ˆ๋‹ค. ๊ฐ Agent๋Š” ์ž์‹ ์˜ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ–ˆ์ง€๋งŒ ์ตœ์ข… ๊ฒฐ๊ณผ๋Š” ์ž˜๋ชป๋์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ Agent๊ฐ€ ํ˜‘๋ ฅํ•˜๋Š” ๊ตฌ์กฐ์—์„œ๋Š” ์‹คํŒจ ์›์ธ์„ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ๋ถ€ํ„ฐ ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. 2/ Execution Responsibility vs Decision Responsibility (EN) The agent executing a transaction is not necessarily the agent responsible for the incorrect decision. An execution agent may faithfully follow instructions generated by a risk assessment agent. Meanwhile, the risk assessment itself may rely on inaccurate information from another source. Simply identifying the wallet that submitted the transaction cannot explain the entire decision process. We need to distinguish who supplied information, who interpreted it, and who authorized execution. (KR) Transaction์„ ์‹คํ–‰ํ•œ Agent๊ฐ€ ๋ฐ˜๋“œ์‹œ ์ž˜๋ชป๋œ ํŒ๋‹จ์˜ ์ฑ…์ž„์ž๋ผ๊ณ  ๋ณผ ์ˆ˜๋Š” ์—†์Šต๋‹ˆ๋‹ค. Execution Agent๋Š” Risk Assessment Agent๊ฐ€ ์ „๋‹ฌํ•œ ๋ช…๋ น์„ ์ •ํ™•ํ•˜๊ฒŒ ์ˆ˜ํ–‰ํ–ˆ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ Risk Assessment ์ž์ฒด๊ฐ€ ๋‹ค๋ฅธ Agent์˜ ๋ถ€์ •ํ™•ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ด๋ฃจ์–ด์กŒ์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Transaction์„ ์ œ์ถœํ•œ Wallet๋งŒ ํ™•์ธํ•ด์„œ๋Š” ์ „์ฒด ์˜์‚ฌ๊ฒฐ์ • ๊ณผ์ •์„ ์„ค๋ช…ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ •๋ณด ์ œ๊ณต, ํŒ๋‹จ, ์‹คํ–‰ ์Šน์ธ์ด๋ผ๋Š” ์—ญํ• ์„ ๊ฐ๊ฐ ๊ตฌ๋ถ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 3/ Reconstructing the Decision Chain (EN) Suppose Agent A provides an incorrect collateral price. Agent B uses that price to calculate a liquidation threshold, and Agent C executes the liquidation. To investigate the incident, we would need to reconstruct the decision chain. What information did Agent B receive? Was the data marked as uncertain? Did Agent C verify the authorization conditions? Transaction logs alone may not capture every off-chain decision. Applications need reliable records connecting inputs, decisions, permissions, and execution results. (KR) Agent A๊ฐ€ ์ž˜๋ชป๋œ ๋‹ด๋ณด ๊ฐ€๊ฒฉ์„ ์ œ๊ณตํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. Agent B๋Š” ํ•ด๋‹น ๊ฐ€๊ฒฉ์œผ๋กœ Liquidation Threshold๋ฅผ ๊ณ„์‚ฐํ•˜๊ณ , Agent C๋Š” ์ฒญ์‚ฐ์„ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ๊ณ  ์›์ธ์„ ํ™•์ธํ•˜๋ ค๋ฉด ์ „์ฒด Decision Chain์„ ์žฌ๊ตฌ์„ฑํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Agent B๋Š” ์–ด๋–ค ์ •๋ณด๋ฅผ ๋ฐ›์•˜๋Š”์ง€, ๋ฐ์ดํ„ฐ์˜ ๋ถˆํ™•์‹ค์„ฑ์„ ์ธ์ง€ํ–ˆ๋Š”์ง€, Agent C๋Š” ์‹คํ–‰ ๊ถŒํ•œ์„ ๊ฒ€์ฆํ–ˆ๋Š”์ง€ ํ™•์ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. On-chain Transaction Log๋งŒ์œผ๋กœ๋Š” ๋ชจ๋“  Off-chain ํŒ๋‹จ์„ ํŒŒ์•…ํ•˜๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Input, Decision, Permission, Execution Result๋ฅผ ์—ฐ๊ฒฐํ•˜๋Š” ๊ธฐ๋ก์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 4/ When Multiple Agents Share Responsibility (EN) Not every failure has a single cause. Agent A might provide outdated data, while Agent B ignores a freshness warning. Agent C could then execute the transaction without checking an additional safety condition. In this situation, assigning all responsibility to one agent would overlook the other failures. A useful accountability framework should distinguish the original error from failures that allowed the error to propagate. However, technical fault attribution does not automatically determine legal or financial liability. (KR) ๋ชจ๋“  ์‚ฌ๊ณ ์˜ ์›์ธ์ด ํ•˜๋‚˜์ธ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. Agent A๊ฐ€ ์˜ค๋ž˜๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ–ˆ๊ณ , Agent B๋Š” ๋ฐ์ดํ„ฐ ๊ฐฑ์‹  ๊ฒฝ๊ณ ๋ฅผ ๋ฌด์‹œํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. Agent C ์—ญ์‹œ ์ถ”๊ฐ€์ ์ธ Safety Condition์„ ํ™•์ธํ•˜์ง€ ์•Š๊ณ  Transaction์„ ์‹คํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ ํ•˜๋‚˜์˜ Agent์— ๋ชจ๋“  ์ฑ…์ž„์„ ๋ถ€์—ฌํ•˜๋ฉด ๋‹ค๋ฅธ ๋ฌธ์ œ๋“ค์„ ๋†“์น˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ตœ์ดˆ ์˜ค๋ฅ˜์™€ ์˜ค๋ฅ˜๊ฐ€ ํ™•์‚ฐ๋˜๋Š” ๊ฒƒ์„ ๋ง‰์ง€ ๋ชปํ•œ ์‹คํŒจ๋ฅผ ๊ตฌ๋ถ„ํ•  ํ•„์š”๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ ๊ธฐ์ˆ ์ ์ธ ์˜ค๋ฅ˜ ๊ธฐ์—ฌ๋„๋ฅผ ํ™•์ธํ•˜๋Š” ๊ฒƒ๊ณผ ๋ฒ•์ ยท๊ธˆ์ „์  ์ฑ…์ž„์„ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. 5/ Preventing the Blame Game (EN) Without verifiable records, agents could produce conflicting explanations after an incident. One agent might claim that another supplied incorrect data. Another might argue that its instructions were misunderstood. This creates a difficult problem when autonomous systems operate without continuous human supervision. Signed messages, versioned inputs, authorization records, and reproducible execution traces could help establish what actually happened. The objective is not to make every agent liable, but to prevent responsibility from disappearing between them. (KR) ๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ๊ธฐ๋ก์ด ์—†๋‹ค๋ฉด ์‚ฌ๊ณ  ๋ฐœ์ƒ ํ›„ Agent๋“ค์ด ์„œ๋กœ ๋‹ค๋ฅธ ์„ค๋ช…์„ ๋‚ด๋†“์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•œ Agent๋Š” ์ž˜๋ชป๋œ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ›์•˜๋‹ค๊ณ  ์ฃผ์žฅํ•˜๊ณ , ๋‹ค๋ฅธ Agent๋Š” ์ž์‹ ์˜ ์ง€์‹œ๊ฐ€ ์ž˜๋ชป ํ•ด์„๋๋‹ค๊ณ  ์ฃผ์žฅํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‚ฌ๋žŒ์ด ์ง€์†์ ์œผ๋กœ ๊ฐ๋…ํ•˜์ง€ ์•Š๋Š” Autonomous System์—์„œ๋Š” ์ด๋Ÿฐ ๋ฌธ์ œ๊ฐ€ ๋”์šฑ ๋ณต์žกํ•ด์ง‘๋‹ˆ๋‹ค. Signed Message, Versioned Input, Authorization Record, Execution Trace ๋“ฑ์„ ํ™œ์šฉํ•˜๋ฉด ์‹ค์ œ ๊ณผ์ •์„ ๊ฒ€์ฆํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ํ•ต์‹ฌ์€ ๋ชจ๋“  Agent์—๊ฒŒ ์ฑ…์ž„์„ ๋ถ€๊ณผํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ Agent ์‚ฌ์ด์—์„œ ์ฑ…์ž„์˜ ๊ทผ๊ฑฐ๊ฐ€ ์‚ฌ๋ผ์ง€์ง€ ์•Š๋„๋ก ๋งŒ๋“œ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. 6/ Rialo and Verifiable Agent Coordination (EN) Rialo's execution environment provides an interesting foundation for exploring these problems. Reactive Transactions could support predefined responses when suspicious conditions are detected. REX and confidential execution could also be relevant to workflows involving sensitive agent information. However, these capabilities do not automatically establish a complete accountability framework. Applications would still need to define how agent identities, decision records, permissions, and dispute procedures are managed. Verifiable execution is an important building block, but accountability requires more than execution alone. (KR) Rialo์˜ Execution Environment๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํƒ๊ตฌํ•  ์ˆ˜ ์žˆ๋Š” ํฅ๋ฏธ๋กœ์šด ๊ธฐ๋ฐ˜์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Reactive Transactions๋ฅผ ํ™œ์šฉํ•˜๋ฉด ์ด์ƒ ์กฐ๊ฑด์ด ๊ฐ์ง€๋์„ ๋•Œ ์‚ฌ์ „์— ์ •์˜๋œ ๋Œ€์‘ ์ ˆ์ฐจ๋ฅผ ์‹คํ–‰ํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. REX์™€ Confidential Execution ์—ญ์‹œ ๋ฏผ๊ฐํ•œ Agent ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•˜๋Š” Workflow์—์„œ ํ™œ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด๋Ÿฌํ•œ ๊ธฐ๋Šฅ๋งŒ์œผ๋กœ ์™„์ „ํ•œ Accountability Framework๊ฐ€ ์ž๋™์œผ๋กœ ๋งŒ๋“ค์–ด์ง€๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. Agent Identity, Decision Record, Permission, Dispute Procedure ๋“ฑ์„ Application ์ฐจ์›์—์„œ ์ •์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ์‹คํ–‰์€ ์ค‘์š”ํ•œ ์ถœ๋ฐœ์ ์ด์ง€๋งŒ ์ฑ…์ž„ ๊ตฌ์กฐ๋ฅผ ์™„์„ฑํ•˜๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a multi-agent lending demonstration involving three independent agents. The first supplies market data, the second evaluates risk, and the third executes transactions. An intentionally outdated price causes an incorrect liquidation. The demonstration should then reconstruct the decision chain and identify whether the failure originated from bad data, ignored warnings, or insufficient execution safeguards. Ideally, the system would also demonstrate how verified evidence can initiate a controlled remediation process. This would make agent accountability a practical engineering problem rather than an abstract promise. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ ๊ฒƒ์€ ์„ธ ๊ฐœ์˜ ๋…๋ฆฝ์ ์ธ Agent๊ฐ€ ์ฐธ์—ฌํ•˜๋Š” Lending Demo์ž…๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ Agent๋Š” Market Data๋ฅผ ์ œ๊ณตํ•˜๊ณ , ๋‘ ๋ฒˆ์งธ๋Š” Risk๋ฅผ ํ‰๊ฐ€ํ•˜๋ฉฐ, ์„ธ ๋ฒˆ์งธ๋Š” Transaction์„ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ์˜๋„์ ์œผ๋กœ ์˜ค๋ž˜๋œ ๊ฐ€๊ฒฉ ์ •๋ณด๋ฅผ ์ „๋‹ฌํ•ด ์ž˜๋ชป๋œ Liquidation์„ ๋ฐœ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ดํ›„ Decision Chain์„ ์žฌ๊ตฌ์„ฑํ•ด ๋ฐ์ดํ„ฐ ์˜ค๋ฅ˜์ธ์ง€, ๊ฒฝ๊ณ  ๋ฌด์‹œ์ธ์ง€, ์‹คํ–‰ ๋‹จ๊ณ„์˜ ์•ˆ์ „์žฅ์น˜ ๋ถ€์กฑ์ธ์ง€ ํ™•์ธํ•˜๋Š” ๊ณผ์ •์„ ๋ณด๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ๋‚˜์•„๊ฐ€ ๊ฒ€์ฆ๋œ Evidence๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ Remediation Process๋ฅผ ์‹œ์ž‘ํ•˜๋Š” ๋ชจ์Šต๊นŒ์ง€ ๋ณด์—ฌ์ค€๋‹ค๋ฉด ๋”์šฑ ์˜๋ฏธ ์žˆ์„ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. Agent Accountability๊ฐ€ ์ถ”์ƒ์ ์ธ ๊ฐœ๋…์ด ์•„๋‹ˆ๋ผ ์‹ค์ œ๋กœ ๊ตฌํ˜„ํ•˜๊ณ  ๊ฒ€์ฆํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ์ˆ ์  ๊ณผ์ œ๋ผ๋Š” ์ ์„ ๋ณด์—ฌ์ค„ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. (EN) As autonomous finance evolves, multi-agent collaboration will make financial decisions increasingly complex. A transaction may have one executor, but its outcome can depend on information and decisions produced by several independent agents. The challenge is not simply identifying who executed a transaction. It is preserving enough verifiable evidence to understand how the decision was made, where safeguards failed, and what corrective action is justified. Autonomous systems should not only be capable of acting independently. They should also make their failures accountable. (KR) Autonomous Finance๊ฐ€ ๋ฐœ์ „ํ• ์ˆ˜๋ก ์—ฌ๋Ÿฌ Agent์˜ ํ˜‘์—…์œผ๋กœ ์ด๋ฃจ์–ด์ง€๋Š” ๊ธˆ์œต ์˜์‚ฌ๊ฒฐ์ •์€ ๋”์šฑ ๋ณต์žกํ•ด์งˆ ๊ฒƒ์ž…๋‹ˆ๋‹ค. Transaction์„ ์‹คํ–‰ํ•œ ์ฃผ์ฒด๋Š” ํ•˜๋‚˜์ผ ์ˆ˜ ์žˆ์ง€๋งŒ, ๊ทธ ๊ฒฐ๊ณผ์—๋Š” ์—ฌ๋Ÿฌ ๋…๋ฆฝ์ ์ธ Agent๊ฐ€ ์ œ๊ณตํ•œ ์ •๋ณด์™€ ํŒ๋‹จ์ด ๋ฐ˜์˜๋ฉ๋‹ˆ๋‹ค. ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋ˆ„๊ฐ€ Transaction์„ ์‹คํ–‰ํ–ˆ๋Š”์ง€๋งŒ ํ™•์ธํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ์–ด๋–ค ๊ณผ์ •์„ ๊ฑฐ์ณ ๊ฒฐ์ •์ด ๋‚ด๋ ค์กŒ๊ณ , ์–ด๋””์—์„œ ์•ˆ์ „์žฅ์น˜๊ฐ€ ์ž‘๋™ํ•˜์ง€ ์•Š์•˜์œผ๋ฉฐ, ์–ด๋–ค ํ›„์† ์กฐ์น˜๊ฐ€ ์ •๋‹นํ•œ์ง€ ๊ฒ€์ฆํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Autonomous System์€ ์Šค์Šค๋กœ ํ–‰๋™ํ•˜๋Š” ๋Šฅ๋ ฅ๋ฟ ์•„๋‹ˆ๋ผ ์‹คํŒจํ–ˆ์„ ๋•Œ ๊ทธ ์›์ธ๊ณผ ์ฑ…์ž„์„ ์ถ”์ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์กฐ๊นŒ์ง€ ๊ฐ–์ถฐ์•ผ ํ•˜์ง€ ์•Š์„๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
๐Ÿ“š Rialo Research Series 62: Financial Remediation - Can Autonomous Finance Repair the Damage? (Kor : ๊ธˆ์œต ํ”ผํ•ด ๋ณต๊ตฌ - ์ž์œจ ๊ธˆ์œต ์‹œ์Šคํ…œ์€ ์ž˜๋ชป๋œ ๊ฑฐ๋ž˜์˜ ํ”ผํ•ด๊นŒ์ง€ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์„๊นŒ?) 1/ Intro: Correcting Trust Doesn't Reverse Losses (EN) In the previous chapter, we explored how an agent might recover from manipulated confidence scores. But restoring trust doesn't undo financial damage. Imagine a lending protocol liquidating a healthy position because corrupted oracle data made its collateral appear insufficient. The oracle is later corrected, but the collateral has already been sold. Fixing the decision model and repairing its consequences are separate problems. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ์กฐ์ž‘๋œ Confidence Score๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์‹ ๋ขฐ๋„๋ฅผ ๋ฐ”๋กœ์žก๋Š”๋‹ค๊ณ  ์ด๋ฏธ ๋ฐœ์ƒํ•œ ๊ธˆ์œต ์†์‹ค๊นŒ์ง€ ์‚ฌ๋ผ์ง€๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ž˜๋ชป๋œ Oracle ๊ฐ€๊ฒฉ ๋•Œ๋ฌธ์— ์ •์ƒ์ ์ธ ๋‹ด๋ณด ํฌ์ง€์…˜์ด Liquidation๋๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ๊ฐ€๊ฒฉ ์ •๋ณด๊ฐ€ ์ˆ˜์ •๋˜๋”๋ผ๋„ ๋‹ด๋ณด๋Š” ์ด๋ฏธ ๋งค๊ฐ๋์Šต๋‹ˆ๋‹ค. ํŒ๋‹จ ๊ธฐ์ค€์„ ๋ณต๊ตฌํ•˜๋Š” ๊ฒƒ๊ณผ ์‹ค์ œ ํ”ผํ•ด๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. 2/ Why Transactions Cannot Simply Be Reversed (EN) Once a blockchain transaction is finalized, an application generally cannot erase it as though it never happened. A liquidation may involve collateral transfers, debt repayment, and third-party buyers. Reversing one operation could require undoing several dependent transactions. Financial remediation therefore often means new compensating actions rather than rewriting history. (KR) Blockchain Transaction์ด Finalized๋˜๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ์‹คํ–‰ ๊ธฐ๋ก์„ ์—†์—ˆ๋˜ ์ผ์ฒ˜๋Ÿผ ๋˜๋Œ๋ฆด ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. Liquidation ๊ณผ์ •์—๋Š” ๋‹ด๋ณด ์ด์ „, ๋ถ€์ฑ„ ์ƒํ™˜, ์ œ3์ž์˜ ์ž์‚ฐ ๋งค์ž… ๋“ฑ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜๋‚˜์˜ ๊ฑฐ๋ž˜๋ฅผ ๋˜๋Œ๋ฆฌ๋ ค๋ฉด ์—ฐ๊ฒฐ๋œ ๋‹ค๋ฅธ ๊ฑฐ๋ž˜๊นŒ์ง€ ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ”ผํ•ด ๋ณต๊ตฌ๋Š” ๊ณผ๊ฑฐ ๊ธฐ๋ก์„ ์‚ญ์ œํ•˜๊ธฐ๋ณด๋‹ค ์ƒˆ๋กœ์šด ๋ณด์ƒ Transaction์„ ์‹คํ–‰ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ ‘๊ทผํ•ด์•ผ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 3/ Compensating Transactions (EN) Suppose a user loses $5,000 through an incorrect liquidation. A protocol could compensate the user from a designated reserve, subject to predefined eligibility and approval rules. But compensation is not identical to reversal. Market prices may have changed, and the original assets may no longer be available. The policy must define what is compensated and how losses are calculated. (KR) ์ž˜๋ชป๋œ Liquidation์œผ๋กœ ์‚ฌ์šฉ์ž๊ฐ€ 5,000๋‹ฌ๋Ÿฌ์˜ ์†์‹ค์„ ์ž…์—ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. Protocol์€ ์‚ฌ์ „์— ๋งˆ๋ จํ•œ Reserve๋ฅผ ํ™œ์šฉํ•ด ๋ณด์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ ๋ณด์ƒ ๋Œ€์ƒ๊ณผ ์Šน์ธ ์กฐ๊ฑด์€ ๋ฏธ๋ฆฌ ์ •์˜๋ผ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ณด์ƒ์€ ์›์ƒ๋ณต๊ตฌ์™€ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์ž์‚ฐ ๊ฐ€๊ฒฉ์ด ๋ณ€ํ–ˆ๊ฑฐ๋‚˜ ๊ธฐ์กด ๋‹ด๋ณด๋ฅผ ๋‹ค์‹œ ํ™•๋ณดํ•  ์ˆ˜ ์—†์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์–ด๋–ค ์†์‹ค์„ ์–ด๋–ค ๊ธฐ์ค€์œผ๋กœ ๋ณด์ƒํ• ์ง€ ๋ช…ํ™•ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 4/ The Risk of False Compensation Claims (EN) Remediation mechanisms can become attack targets. An attacker might deliberately create suspicious transactions and claim they were harmed by an oracle failure. If compensation relies only on an agent's explanation, fraudulent claims may be difficult to distinguish from genuine losses. Eligibility should be checked against historical state, execution records, and incident-specific evidence. (KR) ํ”ผํ•ด ๋ณด์ƒ ๊ธฐ๋Šฅ ์ž์ฒด๊ฐ€ ๊ณต๊ฒฉ ๋Œ€์ƒ์ด ๋  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์ž๊ฐ€ ์˜๋„์ ์œผ๋กœ ๋น„์ •์ƒ์ ์ธ ๊ฑฐ๋ž˜๋ฅผ ๋งŒ๋“  ๋’ค Oracle ์žฅ์• ๋กœ ์†ํ•ด๋ฅผ ์ž…์—ˆ๋‹ค๊ณ  ์ฃผ์žฅํ•˜๋Š” ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. Agent์˜ ์„ค๋ช…๋งŒ์œผ๋กœ ๋ณด์ƒํ•œ๋‹ค๋ฉด ํ—ˆ์œ„ ์ฒญ๊ตฌ๋ฅผ ๊ตฌ๋ถ„ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Historical State, Transaction Record, ์žฅ์•  ๋‹น์‹œ์˜ Evidence๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๋ณด์ƒ ์ž๊ฒฉ์„ ๊ฒ€์ฆํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 5/ Who Pays for the Damage? (EN) Even when an incorrect liquidation is proven, responsibility may remain disputed. Was the loss caused by a faulty oracle, an application bug, or an agent using outdated data? Possible arrangements include insurance reserves, protocol-funded compensation, or contractual liability. None automatically follows from the execution record. Responsibility must be established through governance, agreements, and applicable rules. (KR) ์ž˜๋ชป๋œ Liquidation์ด ํ™•์ธ๋ผ๋„ ๋ˆ„๊ฐ€ ์†์‹ค์„ ๋ถ€๋‹ดํ• ์ง€๋Š” ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. Oracle ์˜ค๋ฅ˜์ธ์ง€, Smart Contract Bug์ธ์ง€, Agent๊ฐ€ ์˜ค๋ž˜๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ ๊ฒƒ์ธ์ง€์— ๋”ฐ๋ผ ์ฑ…์ž„ ๊ตฌ์กฐ๊ฐ€ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Insurance Reserve๋‚˜ Protocol Fund, ์‚ฌ์ „์— ์ •์˜๋œ ๊ณ„์•ฝ์ƒ ์ฑ…์ž„ ๋“ฑ์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ–‰ ๊ธฐ๋ก์ด ์›์ธ์„ ๋ณด์—ฌ์ฃผ๋”๋ผ๋„ ๋ณด์ƒ ์ฑ…์ž„๊นŒ์ง€ ์ž๋™์œผ๋กœ ๊ฒฐ์ •ํ•ด ์ฃผ์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. 6/ Rialo and Automated Remediation (EN) Rialo's Reactive Transactions model supports condition-based transaction triggering during block execution. Applications could use reactive workflows to detect qualifying incidents and initiate predefined remediation procedures. However, reactive execution does not automatically reverse finalized transactions or determine compensation liability. Loss calculations, approvals, funding, and dispute resolution remain application-level responsibilities. (KR) Rialo์˜ Reactive Transactions๋ฅผ ํ™œ์šฉํ•˜๋ฉด ํŠน์ • ์žฅ์•  ์กฐ๊ฑด์ด ํ™•์ธ๋์„ ๋•Œ ์‚ฌ์ „์— ์ •์˜๋œ Remediation Workflow๋ฅผ ์‹œ์ž‘ํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ์„ค๊ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Reactive Execution์ด Finalized Transaction์„ ์ž๋™์œผ๋กœ ์ทจ์†Œํ•˜๊ฑฐ๋‚˜ ๋ณด์ƒ ์ฑ…์ž„์„ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์†์‹ค ๊ณ„์‚ฐ, ๋ณด์ƒ ์Šน์ธ, ์žฌ์› ํ™•๋ณด, ๋ถ„์Ÿ ํ•ด๊ฒฐ์€ Application์—์„œ ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ž๋™ ์‹คํ–‰๊ณผ ์ž๋™ ๋ณด์ƒ์€ ๊ฐ™์€ ๊ฐœ๋…์ด ์•„๋‹™๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a reactive lending demonstration where manipulated oracle data causes an incorrect liquidation. After detecting the error, the application should reconstruct the historical state, calculate the verified loss, and process a controlled compensation request. The demonstration should also reject a fraudulent claim based on unrelated transactions. This would show how verifiable evidence can support remediation without promising impossible transaction reversals. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ Demo๋Š” ์กฐ์ž‘๋œ Oracle ๋ฐ์ดํ„ฐ๋กœ ์ž˜๋ชป๋œ Liquidation์ด ๋ฐœ์ƒํ•˜๋Š” ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. ์˜ค๋ฅ˜ ๋ฐœ๊ฒฌ ํ›„ Historical State๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๊ณ  ์‹ค์ œ ์†์‹ค์„ ๊ณ„์‚ฐํ•œ ๋’ค ๊ฒ€์ฆ๋œ ๋ณด์ƒ ์š”์ฒญ์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ณผ์ •์„ ํ™•์ธํ•˜๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์— ๋ฌด๊ด€ํ•œ ๊ฑฐ๋ž˜๋ฅผ ์ด์šฉํ•œ ํ—ˆ์œ„ ๋ณด์ƒ ์ฒญ๊ตฌ๋Š” ๊ฑฐ๋ถ€๋ผ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํ”ผํ•ด ๋ณต๊ตฌ์˜ ํ•ต์‹ฌ์€ ๊ฑฐ๋ž˜๋ฅผ ๋˜๋Œ๋ฆฌ๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ๊ทผ๊ฑฐ๋กœ ์†์‹ค์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. (EN) We've explored why financial remediation requires evidence, compensation rules, and clear responsibility. But if several agents contribute to one incorrect decision, how should an autonomous system divide accountability among them? (KR) ์ด๋ฒˆ ํŽธ์—์„œ๋Š” Financial Remediation์„ ์ค‘์‹ฌ์œผ๋กœ ์ž˜๋ชป๋œ ์ž๋™ํ™” ํŒ๋‹จ์ด ์‹ค์ œ ๊ธˆ์œต ์†์‹ค๋กœ ์ด์–ด์กŒ์„ ๋•Œ ํ•„์š”ํ•œ ๋ณด์ƒ๊ณผ ๊ฒ€์ฆ ์ ˆ์ฐจ๋ฅผ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์—ฌ๋Ÿฌ Agent๊ฐ€ ํ•˜๋‚˜์˜ ์ž˜๋ชป๋œ ๊ฒฐ์ •์— ๊ด€์—ฌํ–ˆ๋‹ค๋ฉด ์ฑ…์ž„์€ ์–ด๋–ป๊ฒŒ ๋‚˜๋ˆ ์•ผ ํ• ๊นŒ์š”? Oracle ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•œ Agent, ์œ„ํ—˜์„ ํ‰๊ฐ€ํ•œ Agent, Transaction์„ ์‹คํ–‰ํ•œ Agent๊ฐ€ ๋ชจ๋‘ ๋‹ค๋ฅธ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ–ˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Autonomous Finance์—์„œ๋Š” ์‹คํ–‰ ๊ณผ์ •์˜ ์—ญํ• ๋ฟ ์•„๋‹ˆ๋ผ ์‹คํŒจ์— ๋Œ€ํ•œ ์ฑ…์ž„๊นŒ์ง€ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•˜์ง€ ์•Š์„๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
1
1
150
๋ฏธ๊ตญ ์˜ˆ์ธก์‹œ์žฅ ๊ทœ์ œ, ํ•˜์› ๊ฐ๋…์œ„์›ํšŒ ๋ฉ”๋ชจ์™€ CFTC ์œ„์›์žฅ ๋ฐœ์–ธ์œผ๋กœ ๋‹ค์‹œ ๋ณด๋Š” ์Ÿ์  ํ•œ๊ตญ์‹œ๊ฐ„ ์–ด์ ฏ๋ฐค๋ถ€ํ„ฐ ์˜ค๋Š˜ ์ƒˆ๋ฒฝ ์‚ฌ์ด ๋ฏธ๊ตญ์—์„œ ์˜ˆ์ธก์‹œ์žฅ์„ ๋‘˜๋Ÿฌ์‹ผ ์†Œ์‹ ๋‘ ๊ฐœ๊ฐ€ ์ž‡๋”ฐ๋ผ ๋‚˜์™”๋‹ค. ํ•˜์› ๊ฐ๋…ยท์ •๋ถ€๊ฐœํ˜์œ„์›ํšŒ๊ฐ€ ์˜ˆ์ธก์‹œ์žฅ ๋‚ด๋ถ€์ž ๊ฑฐ๋ž˜ ์กฐ์‚ฌ ๊ฒฐ๊ณผ๋ฅผ ๋‹ด์€ ์ง์› ๋ฉ”๋ชจ๋ฅผ ๊ณต๊ฐœํ–ˆ๊ณ  ๋งˆ์ดํด ์…€๋ฆฌ๊ทธ CFTC ์œ„์›์žฅ์€ ํ•œ ์ฝ˜ํผ๋Ÿฐ์Šค์—์„œ ์˜ˆ์ธก์‹œ์žฅ ์—…์ฒด ๊ฒฝ์˜์ง„์„ ๊ฒจ๋ƒฅํ•œ ์ฃผ ์ •๋ถ€๋“ค์˜ ํ˜•์‚ฌ ๋Œ€์‘์„ ๊ฐ•ํ•˜๊ฒŒ ๋น„ํŒํ–ˆ๋‹ค. ๋‘ ์†Œ์‹์€ ๋ฐฉํ–ฅ์ด ์กฐ๊ธˆ ๋‹ค๋ฅธ๋ฐ ์˜ํšŒ๋Š” ๊ณต์ง์ž๊ฐ€ ์˜ˆ์ธก์‹œ์žฅ์—์„œ ๊ฑฐ๋ž˜ํ•˜๋Š” ๊ฒƒ ์ž์ฒด๋ฅผ ๋” ๋ง‰์ž๊ณ  ํ•˜๊ณ  CFTC๋Š” ์˜ˆ์ธก์‹œ์žฅ์„ ์—ฐ๋ฐฉ ํŒŒ์ƒ์ƒํ’ˆ์œผ๋กœ ์ง€ํ‚ค๋ ค ํ•œ๋‹ค. ์นผ์‹œ์™€ ํด๋ฆฌ๋งˆ์ผ“ ๊ฐ™์€ ํ”Œ๋žซํผ์€ ๊ทธ ์‚ฌ์ด์—์„œ ๊ทœ์ œ ๋ฒ”์œ„๊ฐ€ ์ •ํ•ด์ง€๊ธฐ๋ฅผ ๊ธฐ๋‹ค๋ฆฌ๋Š” ์ฒ˜์ง€์ด๋‹ค. ํฐ ํ‹€์€ ์•„์ง ์ •ํ•ด์ง€์ง€ ์•Š์•˜๋‹ค. ํ•˜์› ๋ฉ”๋ชจ๊ฐ€ ์ •๋ฆฌํ•œ ๊ฒƒ ๋ฉ”๋ชจ์— ๋”ฐ๋ฅด๋ฉด ์œ„์›ํšŒ๋Š” ์˜ฌํ•ด 5์›” ์นผ์‹œ์™€ ํด๋ฆฌ๋งˆ์ผ“์— ์ž๋ฃŒ๋ฅผ ์š”์ฒญํ•˜๋ฉด์„œ ์กฐ์‚ฌ๋ฅผ ์‹œ์ž‘ํ–ˆ๊ณ  9์›” ๋ง์—๋Š” ํ•˜์ดํผ๋ฆฌํ€ด๋“œ ๋žฉ์Šค์™€ ํฌ๋ฆฝํ† ๋‹ท์ปด, ํ”„๋ฆฌ๋”•ํŠธ์ž‡๊นŒ์ง€ ๋Œ€์ƒ์„ ๋„“ํ˜”๋‹ค. ๊ทธ๋™์•ˆ ์ฒœ์น ๋ฐฑ ์ชฝ ๊ฐ€๊นŒ์šด ๋ฌธ์„œ๋ฅผ ๋ฐ›์•˜๊ณ  ํ”Œ๋žซํผ ๊ด€๊ณ„์ž๋“ค์—๊ฒŒ์„œ ์—ฌ์•ผ ๊ณต๋™ ์ง์› ๋ธŒ๋ฆฌํ•‘์„ ๋‹ค์„ฏ ์ฐจ๋ก€ ๋“ค์—ˆ๋‹ค. ๋Œ€์ƒ์ด ๋Š˜์—ˆ๊ณ  ๊ธฐ๊ฐ„๋„ ๋ฐ˜๋…„ ๊ฐ€๊นŒ์ด ๊ฑธ๋ ธ๋‹ค. ๊ฒฐ๋ก ์˜ ์ฒซ ์ค„์€ ์ง€๊ธˆ์˜ ๋‚ด๋ถ€์ž ๊ฑฐ๋ž˜ ๋ฒ•๊ทœ๊ฐ€ ์˜ˆ์ธก์‹œ์žฅ์—๋„ ๊ทธ๋Œ€๋กœ ์ ์šฉ๋œ๋‹ค๋Š” ๊ฒƒ์ธ๋ฐ ๋ฉ”๋ชจ๋Š” 2010๋…„ ๊ฐœ์ •๋œ ์ƒํ’ˆ๊ฑฐ๋ž˜๋ฒ•๊ณผ ๊ทธ ์ด๋“ฌํ•ด CFTC๊ฐ€ ๋งŒ๋“  ์‚ฌ๊ธฐยท๊ธฐ๋งŒ ๊ธˆ์ง€ ๊ทœ์ •์„ ๊ทผ๊ฑฐ๋กœ ๋“ค์—ˆ๋‹ค. ํŠนํžˆ ์ƒํ’ˆ๊ฑฐ๋ž˜๋ฒ•์˜ ๊ณต์ง์ž ์กฐํ•ญ์€ ์—ฐ๋ฐฉ ๊ณต๋ฌด์›๊ณผ ์˜์›, ์˜ํšŒ ์ง์›, ๋ฒ•๊ด€ ๋“ฑ์ด ์ง๋ฌด๋กœ ์–ป์€ ๋น„๊ณต๊ฐœ ์ •๋ณด๋กœ ์„ ๋ฌผ์ด๋‚˜ ์˜ต์…˜, ์Šค์™€ํ”„๋ฅผ ๊ฑฐ๋ž˜ํ•˜๋Š” ๊ฒƒ์„ ๊ธˆ์ง€ํ•˜๊ณ  ๊ทธ ์ •๋ณด๋ฅผ ๋‚จ์—๊ฒŒ ๋„˜๊ธฐ๋Š” ๊ฒƒ๋„ ๋ง‰๋Š”๋‹ค. ๋ฉ”๋ชจ์˜ ๊ถŒ๊ณ ๋Š” ํ˜„ํ–‰๋ฒ•๋ณด๋‹ค ํ•œ ๊ฑธ์Œ ๋” ๋‚˜๊ฐ„ ๋‚ด์šฉ์ด๋‹ค. ๋‚ด๋ถ€์ž ์ •๋ณด๋ฅผ ์ผ๋Š”์ง€์™€ ์ƒ๊ด€์—†์ด ์—ฐ๋ฐฉ ๊ณต๋ฌด์›๊ณผ ์˜์›, ์˜ํšŒ ์ง์›์ด ์ฐธ์—ฌํ•  ์ˆ˜ ์žˆ๋Š” ์˜ˆ์ธก์‹œ์žฅ ๊ณ„์•ฝ ์ข…๋ฅ˜๋ฅผ ๋ฒ•์œผ๋กœ ์ œํ•œํ•˜์ž๋Š” ๋‚ด์šฉ์ธ๋ฐ ์ตœ์†Œํ•œ ์ •๋ถ€ ๊ฒฐ์ •์ด๋‚˜ ์ •์น˜ ๊ฒฐ๊ณผ์™€ ๊ด€๋ จ๋œ ๊ณ„์•ฝ์€ ๊ฑฐ๋ž˜ํ•˜์ง€ ๋ชปํ•˜๊ฒŒ ํ•˜์ž๊ณ  ์ ์—ˆ๋‹ค. ์˜ˆ์ธก์‹œ์žฅ ๊ฑฐ๋ž˜๋ฅผ ์•„์˜ˆ ์ „๋ฉด ๊ธˆ์ง€ํ•˜๋Š” ๋ฐฉ์•ˆ๋„ ์„ ํƒ์ง€๋กœ ์˜ฌ๋ ธ๊ณ  ํ”Œ๋žซํผ๊ณผ ๊ทœ์ œ๊ธฐ๊ด€์ด ๋‹จ์†์— ํ•„์š”ํ•œ ์ •๋ณด๋ฅผ ์ฃผ๊ณ ๋ฐ›๊ฒŒ ํ•˜์ž๋Š” ์ œ์•ˆ๋„ ๋ถ™์˜€๋‹ค. ์‹ค์ œ๋กœ ๋ฌธ์ œ๊ฐ€ ๋œ ์‚ฌ๋ก€๋“ค ๋ฉ”๋ชจ๊ฐ€ ๊ทผ๊ฑฐ๋กœ ๋“  ์‚ฌ๋ก€๋Š” ๊ฝค ๊ตฌ์ฒด์ ์ธ๋ฐ ์˜ฌํ•ด 4์›” ๊ธฐ์†Œ๋œ ๋ฏธ ์œก๊ตฐ ์ƒ์‚ฌ๋Š” ๋งˆ๋‘๋กœ ์ „ ๋ฒ ๋„ค์ˆ˜์—˜๋ผ ๋Œ€ํ†ต๋ น์„ ์ฒดํฌํ•œ ๊ตฐ์‚ฌ์ž‘์ „ ๊ด€๋ จ ๊ธฐ๋ฐ€์„ ์ด์šฉํ•ด ํด๋ฆฌ๋งˆ์ผ“์—์„œ ์‚ฌ์‹ญ๋งŒ ๋‹ฌ๋Ÿฌ ๋„˜๋Š” ์ด์ต์„ ๋ƒˆ๋‹ค๋Š” ํ˜์˜๋ฅผ ๋ฐ›๊ณ  ์žˆ๋‹ค. 8์›”์—๋Š” ๋ฐฑ์•…๊ด€ ํ”„๋กฌํ”„ํ„ฐ ๋‹ด๋‹น์ž๊ฐ€ ์—…๋ฌด ์ค‘ ์–ป์€ ์ •๋ณด๋กœ ์นผ์‹œ์—์„œ ๋Œ€ํ†ต๋ น ๋ฐœ์–ธ ๊ด€๋ จ ๊ณ„์•ฝ์„ ๊ฑฐ๋ž˜ํ•œ ์ผ๋กœ CFTC๋กœ๋ถ€ํ„ฐ ์‹ญ์น ๋งŒ ์ด์ฒœ ๋‹ฌ๋Ÿฌ ๋‚ฉ๋ถ€ ๋ช…๋ น์„ ๋ฐ›์•˜๋‹ค. ๋‘˜ ๋‹ค ๊ณต์ง์ž ์‚ฌ๊ฑด์ด๋‹ค. ๊ณต์ง์ž๊ฐ€ ์•„๋‹Œ ์‚ฌ๋ก€๋กœ๋Š” ์Šค์œ„์Šค์—์„œ ์ผํ•˜๋˜ ๊ตฌ๊ธ€ ์—”์ง€๋‹ˆ์–ด ๊ฑด์ด ์žˆ๋‹ค. CFTC๋Š” 5์›” ๊ตฌ๊ธ€์˜ ๋น„๊ณต๊ฐœ ๊ฒ€์ƒ‰ ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•ด ํด๋ฆฌ๋งˆ์ผ“๋‹ท์ปด์—์„œ ์˜ฌํ•ด ๊ฐ€์žฅ ๋งŽ์ด ๊ฒ€์ƒ‰๋œ ๋‹จ์–ด๋ฅผ ๋งžํžˆ๋Š” ๊ณ„์•ฝ์— ๋ฒ ํŒ…ํ•˜๊ณ  ๋ฐฑ๋งŒ ๋‹ฌ๋Ÿฌ ๋„˜๊ฒŒ ๋ฒˆ ํ˜์˜๋กœ ๊ทธ๋ฅผ ์—ฐ๋ฐฉ๋ฒ•์›์— ์ œ์†Œํ–ˆ๋‹ค. ํšŒ์‚ฌ ๋‚ด๋ถ€ ์ •๋ณด๋„ ์–ผ๋งˆ๋“ ์ง€ ์˜ˆ์ธก์‹œ์žฅ์œผ๋กœ ์ƒˆ์–ด ๋‚˜๊ฐˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์„ ๋ณด์—ฌ ์ฃผ๋Š” ์‚ฌ๋ก€์ด๋‹ค. ๋ฉ”๋ชจ๋Š” ํ•˜์ดํผ๋ฆฌํ€ด๋“œ์—์„œ๋„ ๋น„์Šทํ•œ ์˜์‹ฌ ์‚ฌ๋ก€๊ฐ€ ์žˆ์—ˆ๋‹ค๊ณ  ์งš์—ˆ๋‹ค. ์–ธ๋ก  ๋ณด๋„๋ฅผ ์ธ์šฉํ•ด ๋ฏธ๊ตญ ๊ด€์„ธ ์ •์ฑ… ๋ฐœํ‘œ ์ง์ „์— ํฐ ๋ ˆ๋ฒ„๋ฆฌ์ง€ ์ˆ ํฌ์ง€์…˜์ด ์žกํ˜”๋˜ ์ผ์ด ๋‚ด๋ถ€์ž ๊ฑฐ๋ž˜ ํŒจํ„ด๊ณผ ๋‹ฎ์•˜๋‹ค๊ณ  ์ ์€ ๊ฒƒ์ด๋‹ค. ํ”Œ๋žซํผ ๊ฐ์‹œ๋Š” ์–ด๋””๊นŒ์ง€ ์™”๋‚˜ ํฅ๋ฏธ๋กœ์šด ๋ถ€๋ถ„์€ ๋ฉ”๋ชจ๊ฐ€ ํ”Œ๋žซํผ์˜ ์—ญํ• ์„ ๊ฝค ๋†’๊ฒŒ ํ‰๊ฐ€ํ–ˆ๋‹ค๋Š” ์ ์ด๋‹ค. ๋ณด๊ณ ๋œ ๋‚ด๋ถ€์ž ๊ฑฐ๋ž˜ ์‚ฌ๊ฑด ๋Œ€๋ถ€๋ถ„์€ ํ”Œ๋žซํผ์ด ๋จผ์ € ์ฐพ์•„๋‚ด ๊ทœ์ œ๊ธฐ๊ด€๊ณผ ์ˆ˜์‚ฌ๊ธฐ๊ด€์— ์•Œ๋ ธ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ํ”„๋กฌํ”„ํ„ฐ ๋‹ด๋‹น์ž ๊ฑด๋„ ์นผ์‹œ๊ฐ€ ์ˆ˜์ƒํ•œ ๊ฑฐ๋ž˜๋ฅผ ๋ฐœ๊ฒฌํ•ด CFTC์— ์‹ ๊ณ ํ–ˆ๊ณ  ๊ตฌ๊ธ€ ์—”์ง€๋‹ˆ์–ด ๊ฑด์€ ํด๋ฆฌ๋งˆ์ผ“์ด ๋จผ์ € ์ฐพ์•„๋ƒˆ๋‹ค. ํด๋ฆฌ๋งˆ์ผ“์€ ๋ฒ• ์ง‘ํ–‰๊ธฐ๊ด€ ์ถœ์‹  ์ „๋ฌธ๊ฐ€๋ฅผ ์˜์ž…ํ•ด ๊ฐ์‹œํŒ€์„ ๋งก๊ฒผ๊ณ  ์ฒด์ด๋„๋ฆฌ์‹œ์Šค์™€ ํ•จ๊ป˜ ํด๋ฆฌ๊ทธ๋ž˜ํ”„๋ผ๋Š” ๊ฐ์‹œ ์†Œํ”„ํŠธ์›จ์–ด๋ฅผ ๋„์ž…ํ–ˆ๋‹ค. ์†Œ์…œ๋ฏธ๋””์–ด์™€ ์ œ๋ณด, ๋ธ”๋ก์ฒด์ธ ๋ถ„์„, ๋‚ด๋ถ€ ๋ฐ์ดํ„ฐ๋ฅผ ๋ชจ์•„ ์ˆ˜์ƒํ•œ ๊ฑฐ๋ž˜๋ฅผ ๊ฑธ๋Ÿฌ ๋‚ด๊ณ  ์‚ฌ๋žŒ์ด ๋‹ค์‹œ ๊ฒ€ํ† ํ•  ๊ฑฐ๋ž˜๋ฅผ ์ถ”์ฒœํ•˜๋Š” ๊ตฌ์กฐ์ด๋‹ค. ๋ฐ˜๋Œ€๋กœ ํ•˜์ดํผ๋ฆฌํ€ด๋“œ๋Š” ๋ฉ”๋ชจ์—์„œ ์ „ํ˜€ ๋‹ค๋ฅธ ํ‰๊ฐ€๋ฅผ ๋ฐ›์•˜๋‹ค. ๋‹ค์„ฏ ๊ณณ ๊ฐ€์šด๋ฐ ๋„ค ๊ณณ์ด ์ƒ๋‹นํ•œ ์ž๋ฃŒ๋ฅผ ๋‚ด๊ฑฐ๋‚˜ ๋ธŒ๋ฆฌํ•‘์„ ํ–ˆ๋Š”๋ฐ ํ•˜์ดํผ๋ฆฌํ€ด๋“œ๋งŒ ์œ„์›ํšŒ๊ฐ€ ์š”์ฒญํ•œ ์‹ ์› ํ™•์ธ๊ณผ ๊ณ ๊ฐํ™•์ธ ์ •์ฑ… ๊ด€๋ จ ๊ธฐ๋ก์„ ๊ฐ–๊ณ  ์žˆ์ง€ ์•Š๋‹ค๊ณ  ๋‹ตํ–ˆ๋‹ค๊ณ  ๋ฉ”๋ชจ๋Š” ์ ์—ˆ๋‹ค. ๋‹ต๋ณ€์€ ๋กœํŽŒ์„ ํ†ตํ•ด ์ „๋‹ฌ๋๋‹ค. ์˜จ์ฒด์ธ ๊ฑฐ๋ž˜์†Œ๋Š” ๊ฑฐ๋ž˜ ๊ธฐ๋ก์ด ๋ชจ๋‘ ๊ณต๊ฐœ๋˜์ง€๋งŒ ๊ทธ ์ฃผ์†Œ๊ฐ€ ๋ˆ„๊ตฌ์ธ์ง€์— ๋Œ€ํ•œ ๊ธฐ๋ก์€ ๋”ฐ๋กœ ์—†๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ๋‹ค. ์ด๋ฒˆ ๋ฉ”๋ชจ์—์„œ ๊ทธ ์ฐจ์ด๊ฐ€ ๊ทธ๋Œ€๋กœ ๋“œ๋Ÿฌ๋‚œ ์…ˆ์ด๋‹ค. ํด๋ฆฌ๋งˆ์ผ“์ด ๋‘ ๊ฐœ์˜ ํ”Œ๋žซํผ์œผ๋กœ ๋‚˜๋‰˜์–ด ์žˆ๋‹ค๋Š” ์ ๋„ ๋ฉ”๋ชจ์— ์ •๋ฆฌ๋ผ ์žˆ๋Š”๋ฐ ๋ฏธ๊ตญ ๊ณ ๊ฐ์„ ๋ฐ›๋Š” ํด๋ฆฌ๋งˆ์ผ“ US๋Š” CFTC์— ๋“ฑ๋ก๋œ ์ง€์ • ๊ณ„์•ฝ์‹œ์žฅ์ด๊ณ  ๊ณ ๊ฐํ™•์ธ๊ณผ ์ž๊ธˆ์„ธํƒ๋ฐฉ์ง€ ๊ทœ์ •์„ ๋”ฐ๋ฅธ๋‹ค. ๋ฐ˜๋ฉด ํŒŒ๋‚˜๋งˆ์— ๋ณธ์‚ฌ๋ฅผ ๋‘” ํด๋ฆฌ๋งˆ์ผ“๋‹ท์ปด์€ CFTC ๋“ฑ๋ก ์‹œ์žฅ์ด ์•„๋‹ˆ๊ณ  ๋ฏธ๊ตญ์ธ์ด ์“ฐ์ง€ ๋ชปํ•˜๊ฒŒ ๋ง‰๊ณ  ์žˆ์ง€๋งŒ VPN ๊ฐ™์€ ์šฐํšŒ ์ˆ˜๋‹จ ๋•Œ๋ฌธ์— ์ผ๋ถ€ ๋ฏธ๊ตญ์ธ์ด ์ด์šฉํ–ˆ์„ ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ฉ”๋ชจ๋Š” ๋ดค๋‹ค. ์ด๋ฆ„์€ ๊ฐ™์•„๋„ ๊ทœ์ œ๋„ ๋‹ค๋ฅด๊ณ  ๊ณ ๊ฐ๋„ ๋‹ค๋ฅด๋‹ค. CFTC์™€ ์ฃผ ์ •๋ถ€์˜ ์‹ธ์›€ ์…€๋ฆฌ๊ทธ ์œ„์›์žฅ์˜ ๋ฐœ์–ธ์€ ์ด ํ๋ฆ„์˜ ๋‹ค๋ฅธ ์ถ•์ด๋‹ค. ๋”๋ธ”๋ก์— ๋”ฐ๋ฅด๋ฉด ๊ทธ๋Š” ์บ”ํ„ฐ ์ปจ๋ฒ„์ „์Šค ์ฝ˜ํผ๋Ÿฐ์Šค์—์„œ ์ฃผ ์ •๋ถ€๋“ค์ด ์—ฐ๋ฐฉ ์ฒด๊ณ„ ์•„๋ž˜ ํ•ฉ๋ฒ•์ ์ธ ์ƒํ’ˆ์„ ๋‚ด๋†“์€ ํšŒ์‚ฌ์˜ ์ฐฝ์—…์ž์™€ CEO๋ฅผ ํ˜•์‚ฌ ๊ธฐ์†Œ๋กœ ๊ฐ€๋‘๋ ค ํ•œ๋‹ค๊ณ  ๋งํ–ˆ๊ณ  ์ด๋ฅผ ๋งค์šฐ ์ง€๋‚˜์น˜๋‹ค๊ณ  ํ‰๊ฐ€ํ–ˆ๋‹ค. ๋„๋ฐ•์ด๋ƒ ํŒŒ์ƒ์ƒํ’ˆ์ด๋ƒ๋ฅผ ๋‘๊ณ  ์˜๊ฒฌ์ด ๊ฐˆ๋ฆฌ๋Š” ๋ฌธ์ œ๋กœ ๊ฐ์˜ฅ์— ๊ฐˆ ์œ„ํ—˜์„ ์ง€์›Œ์„œ๋Š” ์•ˆ ๋œ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ๋ฐœ์–ธ ์ˆ˜์œ„๊ฐ€ ๊ฝค ๋†’์•˜๋‹ค. ๊ทธ๋Š” ์ฃผ๋งˆ๋‹ค ๋”ฐ๋กœ ์†Œ์†ก์„ ๋ฒŒ์ด๊ธฐ๋ณด๋‹ค ์—ฐ๋ฐฉ๋Œ€๋ฒ•์›์—์„œ ๋ณธ์•ˆ์„ ๋‹คํˆฌ๋Š” ํŽธ์ด ๋‚ซ๋‹ค๋Š” ๋œป๋„ ๋ฐํ˜”๊ณ  CFTC๊ฐ€ ์˜ˆ์ธก์‹œ์žฅ์— ๋Œ€ํ•ด ๊ฐ–๋Š” ๋ฐฐํƒ€์  ๊ด€ํ• ๊ถŒ์„ ๊ณ„์† ์ง€ํ‚ค๊ฒ ๋‹ค๊ณ  ๋ง๋ถ™์˜€๋‹ค. ์ด ๋ฐœ์–ธ์˜ ๋ฐฐ๊ฒฝ์—๋Š” ๊ทธ ์ „๋‚  ๋ฏธ๊ตญ์—์„œ ์ž‡๋”ฐ๋ผ ๋‚˜์˜จ ์†Œ์†ก๋“ค์ด ์žˆ๋‹ค. ๋ฏธ ๋ฒ•๋ฌด๋ถ€์™€ CFTC๋Š” ์นผ์‹œ๋ฅผ ๋ถˆ๋ฒ• ๋„๋ฐ• ํ”Œ๋žซํผ์œผ๋กœ ๋ณด๊ณ  ์†Œ์†ก์„ ๋‚ธ ์•„์นธ์†Œ์ฃผ ๋ฒ•๋ฌด์žฅ๊ด€์„ ์ƒ๋Œ€๋กœ ์ฃผ์˜ ์ง‘ํ–‰์„ ๋ง‰์•„ ๋‹ฌ๋ผ๋Š” ์†Œ์†ก์„ ๋ƒˆ๋‹ค. ๊ฐ™์€ ๋‚  ๋ฃจ์ด์ง€์• ๋‚˜์ฃผ ๋ฒ•๋ฌด์žฅ๊ด€๋„ ์นผ์‹œ์™€ ํด๋ฆฌ๋งˆ์ผ“, ๋กœ๋นˆํ›„๋“œ, ์ฝ”์ธ๋ฒ ์ด์Šค ๋“ฑ์„ ์ƒ๋Œ€๋กœ ์†Œ์†ก์„ ์ œ๊ธฐํ–ˆ๋‹ค. ๋”๋ธ”๋ก์€ CFTC๊ฐ€ ์ง€๊ธˆ ์—ด๋‘ ๊ณณ ์•ˆํŒŽ์˜ ์ฃผ์™€ ๋„์‹œ๋ฅผ ์ƒ๋Œ€๋กœ ์ด๋Ÿฐ ๋‹คํˆผ์„ ๋ฒŒ์ด๊ณ  ์žˆ๋‹ค๊ณ  ์ „ํ–ˆ๋‹ค. ํฌ๋ฆฝํ†  ์ชฝ์—์„œ ๋ณผ ์  ์˜ˆ์ธก์‹œ์žฅ์€ ์ด์ œ ํฌ๋ฆฝํ†  ์—…๊ณ„์˜ ์ฃผ์š” ์‚ฌ์—… ๊ฐ€์šด๋ฐ ํ•˜๋‚˜๊ฐ€ ๋๋Š”๋ฐ ํด๋ฆฌ๋งˆ์ผ“์€ ๋ธ”๋ก์ฒด์ธ ์œ„์—์„œ ์ถœ๋ฐœํ–ˆ๊ณ  ์ฝ”์ธ๋ฒ ์ด์Šค์™€ ๋กœ๋นˆํ›„๋“œ ๊ฐ™์€ ๊ฑฐ๋ž˜ ํ”Œ๋žซํผ๋„ ์†Œ์†ก ๋Œ€์ƒ์— ์ด๋ฆ„์„ ์˜ฌ๋ ธ๋‹ค. ํ•˜์ดํผ๋ฆฌํ€ด๋“œ์ฒ˜๋Ÿผ ์˜จ์ฒด์ธ ํŒŒ์ƒ์ƒํ’ˆ ๊ฑฐ๋ž˜์†Œ๊ฐ€ ์˜ํšŒ ์กฐ์‚ฌ ๋Œ€์ƒ์— ๋“ค์–ด๊ฐ„ ๊ฒƒ๋„ ๊ฐ™์€ ํ๋ฆ„์ด๋‹ค. ๊ทธ๋ž˜์„œ ์ด๋ฒˆ ๋ฉ”๋ชจ๋Š” ๋‘ ๊ฐ€์ง€๋ฅผ ํ•จ๊ป˜ ๋ณด์—ฌ ์ค€๋‹ค. ์˜จ์ฒด์ธ ๊ฑฐ๋ž˜๋Š” ๊ธฐ๋ก์ด ๊ณต๊ฐœ๋ผ ์žˆ์–ด ์˜์‹ฌ ๊ฑฐ๋ž˜๋ฅผ ์ฐพ๊ธฐ ์‰ฝ์ง€๋งŒ ์‹ ์› ๊ธฐ๋ก์ด ์—†์œผ๋ฉด ๊ทธ ๊ฑฐ๋ž˜๋ฅผ ๋ˆ„๊ฐ€ ํ–ˆ๋Š”์ง€ ํ™•์ธํ•˜๋Š” ๋‹จ๊ณ„์—์„œ ๋ง‰ํžŒ๋‹ค. ๋ฏธ๊ตญ ์˜ํšŒ๊ฐ€ ํ”Œ๋žซํผ์— ๊ธฐ๋Œ€ํ•˜๋Š” ๊ฐ์‹œ ์ˆ˜์ค€์ด ๋†’์•„์งˆ์ˆ˜๋ก ๊ณ ๊ฐํ™•์ธ ์ฒด๊ณ„๊ฐ€ ์—†๋Š” ์˜จ์ฒด์ธ ์„œ๋น„์Šค๋Š” ๊ฐ™์€ ์งˆ๋ฌธ์„ ๊ณ„์† ๋ฐ›๊ฒŒ ๋  ์ˆ˜ ์žˆ๋‹ค. ์•ž์œผ๋กœ ์ง€์ผœ๋ณผ ๋Œ€๋ชฉ์€ ํฌ๊ฒŒ ์„ธ ๊ฐ€์ง€ ์ •๋„๋กœ ์ •๋ฆฌํ•  ์ˆ˜ ์žˆ๋‹ค. ๋ฉ”๋ชจ์˜ ๊ถŒ๊ณ ๊ฐ€ ์‹ค์ œ ๋ฒ•์•ˆ์œผ๋กœ ์ด์–ด์ง€๋Š”์ง€, ์•„์นธ์†Œ์™€ ๋ฃจ์ด์ง€์• ๋‚˜ ์†Œ์†ก์—์„œ ๋ฒ•์›์ด ์—ฐ๋ฐฉ ๊ด€ํ• ๊ถŒ์„ ์–ด๋–ป๊ฒŒ ๋ณด๋Š”์ง€, ๊ทธ๋ฆฌ๊ณ  ์…€๋ฆฌ๊ทธ ์œ„์›์žฅ ๋ง๋Œ€๋กœ ์ด ๋ฌธ์ œ๊ฐ€ ์—ฐ๋ฐฉ๋Œ€๋ฒ•์›๊นŒ์ง€ ์˜ฌ๋ผ๊ฐ€๋Š”์ง€์ด๋‹ค. ์–ด๋А ์ชฝ์œผ๋กœ ์ •๋ฆฌ๋˜๋“  ์˜ˆ์ธก์‹œ์žฅ์„ ์“ฐ๋Š” ์‚ฌ๋žŒ์€ ์ž์‹ ์ด ์ด์šฉํ•˜๋Š” ํ”Œ๋žซํผ์ด ์–ด๋А ๋‚˜๋ผ์˜ ์–ด๋–ค ๊ทœ์ œ๋ฅผ ๋ฐ›๋Š” ๊ณณ์ธ์ง€๋ถ€ํ„ฐ ํ™•์ธํ•ด ๋‘๋Š” ํŽธ์ด ๋‚ซ๋‹ค. ์šฐ๋ฆฌ๋‚˜๋ผ๋Š” ์ด๋ฏธ ์ด ๋ฌธ์ œ์— ๋‹ต์„ ๋‚ด๋†“์€ ์ชฝ์— ๊ฐ€๊น๋‹ค. ๋ฐฉ์†ก๋ฏธ๋””์–ดํ†ต์‹ ์‹ฌ์˜์œ„์›ํšŒ๋Š” ์ง€๋‚œ 8์›” 18์ผ ํด๋ฆฌ๋งˆ์ผ“์˜ ์‚ฌํ–‰์„ฑ์„ ์ธ์ •ํ•ด ๊ตญ๋‚ด ์ ‘์†์ฐจ๋‹จ์„ ์˜๊ฒฐํ–ˆ๋Š”๋ฐ, ํ˜•๋ฒ•์ƒ ๋„๋ฐ• ๋ฐฉ์กฐ์™€ ๋„๋ฐ•์žฅ์†Œ ๊ฐœ์„ค์„ ๋ชฉ์ ์œผ๋กœ ํ•˜๋Š” ์ •๋ณด, ๊ตญ๋ฏผ์ฒด์œก์ง„ํฅ๋ฒ•์ƒ ์œ ์‚ฌํ–‰์œ„๋ฅผ ๋ชฉ์ ์œผ๋กœ ํ•˜๋Š” ์ •๋ณด์— ํ•ด๋‹นํ•œ๋‹ค๋Š” ํŒ๋‹จ์ด์—ˆ๋‹ค. ํด๋ฆฌ๋งˆ์ผ“์€ ํ•œ๊ตญ์–ด ์„œ๋น„์Šค๋ฅผ ์—†์•ด๊ณ  ๋น„์ˆ˜ํƒํ˜• ๊ตฌ์กฐ๋ผ ์šด์˜์ž๊ฐ€ ์ฃผ์žฌ์ž์— ํ•ด๋‹นํ•˜์ง€ ์•Š๋Š”๋‹ค๊ณ  ์ฃผ์žฅํ–ˆ์ง€๋งŒ, ์œ„์›ํšŒ๋Š” ๊ธฐ์ˆ ์  ํŠน์„ฑ์ด๋‚˜ ์„œ๋น„์Šค ๋ฐฉ์‹์„ ์ด์œ ๋กœ ๊ตญ๋‚ด ๋ฒ•๋ น ์ ์šฉ์„ ํ”ผํ•  ์ˆ˜๋Š” ์—†๋‹ค๊ณ  ๋ดค๋‹ค. ์ˆ˜์‚ฌ๋„ ์ด๋ฏธ ์ง„ํ–‰ ์ค‘์ด๋‹ค. ๊ฐ•์›๊ฒฝ์ฐฐ์ฒญ์€ 6์›” ๊ตญ๋‚ด ์ด์šฉ์ž ์ˆ˜์‚ฌ์— ์ฐฉ์ˆ˜ํ•œ ๋ฐ ์ด์–ด 9์›”์—๋Š” ์ „๊ตญ์˜ ์ด์šฉ์ž๋“ค์—๊ฒŒ ์ถœ์„์„ ์ผ๊ด„ ํ†ต์ง€ํ–ˆ๋‹ค. ํ˜•๋ฒ• ์ œ246์กฐ๋Š” ๋„๋ฐ•์„ ํ•œ ์‚ฌ๋žŒ์„ 1์ฒœ๋งŒ ์› ์ดํ•˜์˜ ๋ฒŒ๊ธˆ์—, ์ƒ์Šต๋„๋ฐ•์€ 3๋…„ ์ดํ•˜์˜ ์ง•์—ญ ๋˜๋Š” 2์ฒœ๋งŒ ์› ์ดํ•˜์˜ ๋ฒŒ๊ธˆ์— ์ฒ˜ํ•˜๋„๋ก ํ•˜๊ณ  ์žˆ๋‹ค. ๊ทธ๋ž˜์„œ ๋ฏธ๊ตญ์—์„œ ์—ฐ๋ฐฉ๊ณผ ์ฃผ๊ฐ€ ๊ด€ํ• ๊ถŒ์„ ๋‘๊ณ  ๋‹คํˆฌ๋Š” ๊ฒƒ๊ณผ ๋ณ„๊ฐœ๋กœ, ๊ตญ๋‚ด ์ด์šฉ์ž๊ฐ€ ํ•ด์™ธ ์˜ˆ์ธก์‹œ์žฅ์— ๋ˆ์„ ๊ฑฐ๋Š” ์ผ์€ ์ด๋ฏธ ํ˜•์‚ฌ ์ฒ˜๋ฒŒ ์œ„ํ—˜์„ ์•ˆ๊ณ  ์žˆ๋‹ค๊ณ  ๋ด์•ผ ํ•œ๋‹ค. ๊ฒฐ๋ก ์€ ๊ทธ๋ƒฅ ์•ˆํ•˜๋Š”๊ฒŒ ์†ํŽธํ•˜๋‹ค.
1
273
ํŠธ์›ฌํ‹ฐ์› ์บํ”ผํ„ธ์ด ํ•œ๊ตญ์‹œ๊ฐ„ ์–ด์ ฏ๋ฐค ๋ฏธ๊ตญ ์ฆ๊ถŒ๊ฑฐ๋ž˜์œ„์›ํšŒ์— ๋‚ธ 3๋ถ„๊ธฐ ๋ถ„๊ธฐ๋ณด๊ณ ์„œ๋ฅผ ๋ณด๋ฉด ์„ ๋‹ฌ ๋™์•ˆ ์ˆœ์ด์ต์ด ์•ฝ ์‹ญ์–ต ๊ตฌ์ฒœ๋งŒ ๋‹ฌ๋Ÿฌ์˜€๋Š”๋ฐ ๊ฑฐ์˜ ์ „๋ถ€๊ฐ€ ๋ณด์œ ํ•œ ๋น„ํŠธ์ฝ”์ธ์˜ ๊ณต์ •๊ฐ€์น˜ ํ‰๊ฐ€์ด์ต์—์„œ ๋‚˜์™”๋‹ค. ํ…Œ๋”์™€ ๋น„ํŠธํŒŒ์ด๋„ฅ์Šค, ์†Œํ”„ํŠธ๋ฑ…ํฌ์˜ ์ง€์›์œผ๋กœ ์ง€๋‚œํ•ด ๋ด„ ์ถœ๋ฒ”ํ•œ ํšŒ์‚ฌ์ด๋‹ค. ์˜ฌํ•ด 5์›”์—๋Š” ํ…Œ๋”๊ฐ€ ์†Œํ”„ํŠธ๋ฑ…ํฌ ์ง€๋ถ„๊นŒ์ง€ ์‚ฌ๋“ค์˜€๋‹ค. 9์›” ๋ง ๊ธฐ์ค€์œผ๋กœ ๋“ค๊ณ  ์žˆ๋˜ ๋น„ํŠธ์ฝ”์ธ์€ ์‚ฌ๋งŒ ์‚ผ์ฒœ์—ฌ ๊ฐœ์˜€๊ณ  ๊ฐœ๋‹น ํŒ”๋งŒ ์‚ผ์ฒœ์น ๋ฐฑ ๋‹ฌ๋Ÿฌ ์•ˆํŒŽ์œผ๋กœ ํ‰๊ฐ€๋๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ๊ฐ™์€ ์„ ๋‹ฌ ๋™์•ˆ ๋ณธ์—… ์ชฝ ์ˆซ์ž๋Š” ์กฐ์šฉํ–ˆ๋‹ค. ์˜์—…์†์‹ค์ด ์•ฝ ์‚ฌ๋ฐฑํŒ”์‹ญ๋งŒ ๋‹ฌ๋Ÿฌ์˜€๊ณ  ์ด์ž๋น„์šฉ๋„ ๋ฐฑ์‚ผ์‹ญ๋งŒ ๋‹ฌ๋Ÿฌ ๋‚จ์ง“ ๋‚˜๊ฐ”๋‹ค. ์˜ฌํ•ด 1์›”๋ถ€ํ„ฐ 9์›”๊นŒ์ง€ ๋ˆ„์ ์œผ๋กœ ๋ณด๋ฉด ์˜คํžˆ๋ ค ์•ฝ ์ผ์–ต ํŒ”์ฒœ์œก๋ฐฑ๋งŒ ๋‹ฌ๋Ÿฌ ์ˆœ์†์‹ค์ธ๋ฐ ๊ทธ์‚ฌ์ด ๋น„ํŠธ์ฝ”์ธ ๊ฐ€๊ฒฉ์ด ๋‚ด๋ ค๊ฐ„ ์˜ํ–ฅ์ด ๋ฐ˜์˜๋๋‹ค. ๋ผํŒŒ ์ž๊ตฌ๋ฆฌ CEO๋„ ์ฃผ์ฃผ์„œํ•œ์—์„œ ์ด ์ ์„ ๋จผ์ € ์งš์—ˆ๋‹ค. ๋ณด์œ  ์ž์‚ฐ์ด ์˜ฌ๋ผ์„œ ์ƒ๊ธด ์ด์ต๊ณผ ์ง์ ‘ ์šด์˜ํ•˜๋Š” ์‚ฌ์—…์—์„œ ๋ฒˆ ์ด์ต์€ ๊ฐ™์ง€ ์•Š๊ณ  ๋’ค์˜ ๊ฒƒ์„ ๋งŒ๋“œ๋Š” ๊ฒŒ ์ž๊ธฐ ์ผ์ด๋ผ๋Š” ๋‚ด์šฉ์ด๋‹ค. ํšŒ์‚ฌ๋Š” ์ฑ„๊ตด๊ณผ ๋Œ€์ถœ, ํŠธ๋ ˆ์ €๋ฆฌ ์„œ๋น„์Šค, ์—๋„ˆ์ง€ ์ชฝ ๊ธฐํšŒ๋ฅผ ๊ฒ€ํ† ํ•˜๊ณ  ์žˆ๊ณ  ์ฃผ๊ฐ€๊ฐ€ ์ฃผ๋‹น ๊ท€์† ์ž์‚ฐ๊ฐ€์น˜๋ณด๋‹ค ๋‚ฎ๊ฒŒ ๊ฑฐ๋ž˜๋˜๋ฉด ์ž์‚ฌ์ฃผ ๋งค์ž…๋„ ๊ณ ๋ คํ•˜๊ฒ ๋‹ค๊ณ  ๋ฐํ˜”๋‹ค. ๋‹ด๋ณด ๊ตฌ์กฐ๋„ ๊ฐ™์ด ๋ณผ ๋งŒํ•œ๋ฐ ์ด ํšŒ์‚ฌ์˜ ์ „ํ™˜์‚ฌ์ฑ„ ์›๊ธˆ์€ ์•ฝ ์‚ฌ์–ต ํŒ”์ฒœ์œก๋ฐฑ๋งŒ ๋‹ฌ๋Ÿฌ์ด๊ณ  ๊ทธ ๋‹ด๋ณด๋กœ ๋น„ํŠธ์ฝ”์ธ ์•ฝ ๋งŒ ์œก์ฒœ ๊ฐœ๊ฐ€ ์žกํ˜€ ์žˆ๋‹ค. ๋ณด์œ ๋Ÿ‰์˜ ์‚ผ๋ถ„์˜ ์ผ์ด ๋„˜๋Š” ๋ฌผ๋Ÿ‰์ด๋‹ค. ๋น„ํŠธ์ฝ”์ธ์€ ์ง€๊ธˆ ์น ๋งŒ ๋‹ฌ๋Ÿฌ ์•„๋ž˜์—์„œ ์›€์ง์ด๊ณ  ์žˆ๋Š”๋ฐ 9์›” ๋ง ํ‰๊ฐ€ ๊ฐ€๊ฒฉ๋ณด๋‹ค ์ดํ•  ๊ฐ€๊นŒ์ด ๋‚ฎ์€ ์ˆ˜์ค€์ด๋‹ค. ์ฃผ๊ฐ€๋„ ํ•œ ๋‹ฌ ์ƒˆ ์ดํ•  ๊ฐ€๊นŒ์ด ๋น ์กŒ๋‹ค. ๊ณต์ •๊ฐ€์น˜ ํšŒ๊ณ„๋Š” ๊ฐ€๊ฒฉ์ด ์˜ค๋ฅผ ๋•Œ ์ด์ต์„ ํ‚ค์šด ๋งŒํผ ๋‚ด๋ ค๊ฐˆ ๋•Œ๋„ ๊ฐ™์€ ๋ฐฉ์‹์œผ๋กœ ์žฅ๋ถ€์— ๋ฐ˜์˜๋œ๋‹ค. ํŠธ๋ ˆ์ €๋ฆฌ ๊ธฐ์—…์˜ ๋ถ„๊ธฐ ์‹ค์ ์—์„œ ์ˆœ์ด์ต ํ•œ ์ค„๋ณด๋‹ค ์˜์—…์†์ต๊ณผ ๋‹ด๋ณด๋กœ ๋ฌถ์ธ ๋ฌผ๋Ÿ‰์„ ๋จผ์ € ํ™•์ธํ•˜๋Š” ์ด์œ ๊ฐ€ ์—ฌ๊ธฐ์— ์žˆ๋‹ค.
92
๊ณผ๊ฐํ•œ ๊ฒฐ๋‹จ๋ ฅ, ๋„ˆ ์ฐธ ๋Œ€๋‹จํ•˜๋‹ค. @grok
2
2
51
๋ธ”๋ก์ŠคํŠธ๋ฆผ๊ณผ ํƒˆ๋ ˆ์Šค์˜ ๊ธฐ๊ด€์šฉ ์ˆ˜ํƒ, ๊ทธ๋ฆฌ๊ณ  ์• ๋ค ๋ฐฑ์ด ๋‹ค์‹œ ๊บผ๋‚ธ ์ฝœ๋“œ์›”๋ › ์ด์•ผ๊ธฐ ๋ธ”๋ก์ŠคํŠธ๋ฆผ์ด ํ•œ๊ตญ์‹œ๊ฐ„ ์˜ค๋Š˜ ์ƒˆ๋ฒฝ ํ”„๋ž‘์Šค ๋ณด์•ˆ ๊ธฐ์—… ํƒˆ๋ ˆ์Šค์™€ ํ•จ๊ป˜ ๋งŒ๋“  ๊ธฐ๊ด€์šฉ ๋””์ง€ํ„ธ ์ž์‚ฐ ์ˆ˜ํƒ ์†”๋ฃจ์…˜์„ ๊ณต๊ฐœํ–ˆ๋Š”๋ฐ ํ•ต์‹ฌ์€ ์„œ๋ช… ๊ณผ์ • ์ „์ฒด๋ฅผ ํƒˆ๋ ˆ์Šค์˜ ํ•˜๋“œ์›จ์–ด ๋ณด์•ˆ ๋ชจ๋“ˆ(HSM) ์•ˆ์—์„œ ๋Œ๋ฆฐ๋‹ค๋Š” ์ ์ด๋‹ค. ๊ฐ™์€ ๋‚  ๋‰ด์š• ๋””์ง€ํ„ธ ์ž์‚ฐ ํฌ๋Ÿผ ๋ฌด๋Œ€์— ์˜ค๋ฅธ ์• ๋ค ๋ฐฑ CEO๋Š” ์ˆ˜ํƒ์˜ ๊ตํ›ˆ์œผ๋กœ ์ฝœ๋“œ ๋ณด๊ด€์„ ๊ผฝ์•˜๋‹ค. ๊ธฐ๊ด€์šฉ ์ œํ’ˆ ์†Œ์‹์ด์ง€๋งŒ ๋‚ด์šฉ์„ ๋œฏ์–ด๋ณด๋ฉด ๊ฐœ์ธ์—๊ฒŒ๋„ ์“ธ๋ชจ๊ฐ€ ์žˆ๋‹ค. ํฐ ๊ฑฐ๋ž˜์†Œ ํ•ดํ‚น์ด ์ด์–ด์ง„ ์˜ฌํ•ด ๊ฐ€์„์— ํ‚ค๋ฅผ ์–ด๋””์— ๋‘๊ณ  ์–ด๋–ป๊ฒŒ ์„œ๋ช…ํ•˜๋А๋ƒ๋Š” ์งˆ๋ฌธ์€ ๋ˆ„๊ตฌ์—๊ฒŒ๋‚˜ ๊ฐ™์€ ๋ฌด๊ฒŒ๋ฅผ ๊ฐ–๋Š”๋‹ค. ๋ธ”๋ก์ŠคํŠธ๋ฆผ ์„ค๋ช…์— ๋”ฐ๋ฅด๋ฉด ์ด ์†”๋ฃจ์…˜์€ ์ž์‚ฌ์˜ ์‹œํ์–ด MPC ๊ธฐ์ˆ ๊ณผ ํƒˆ๋ ˆ์Šค์˜ HSM์„ ๊ฒฐํ•ฉํ–ˆ๋‹ค. MPC๋Š” ๋‹ค์ž๊ฐ„ ์—ฐ์‚ฐ์˜ ์ค„์ž„๋ง๋กœ ํ•˜๋‚˜์˜ ํ‚ค๋ฅผ ์—ฌ๋Ÿฌ ์กฐ๊ฐ์œผ๋กœ ๋‚˜๋ˆ  ์—ฌ๋Ÿฌ ๋‹น์‚ฌ์ž๊ฐ€ ํ•จ๊ป˜ ์„œ๋ช…ํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ๋ฐฉ์‹์ด๋‹ค. ์–ด๋А ํ•œ ๊ณณ์—๋„ ์˜จ์ „ํ•œ ํ‚ค๊ฐ€ ์—†๋‹ค๋Š” ์ ์ด ์žฅ์ ์ด๋‹ค. ์ผ๋ฐ˜์ ์ธ MPC ๊ตฌํ˜„์ด ์„œ๋ช… ๊ณผ์ •์˜ ์ผ๋ถ€๋ฅผ ์ „์šฉ ๋ณด์•ˆ ํ•˜๋“œ์›จ์–ด ๋ฐ–์—์„œ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ์žˆ๋‹ค. ์ด๋ฒˆ ๊ตฌ์กฐ๋Š” ์„œ๋ช… ํ”„๋กœํ† ์ฝœ ์ „์ฒด๋ฅผ HSM์ด๋ผ๋Š” ๋ณดํ˜ธ๋œ ํ™˜๊ฒฝ ์•ˆ์—์„œ ์‹คํ–‰ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์„œ๋ช…ํ•˜๋Š” ์ˆœ๊ฐ„์—๋„ ํ‚ค ์ž๋ฃŒ๊ฐ€ ํ•˜๋“œ์›จ์–ด ๋ฐ–์œผ๋กœ ๋‚˜์˜ค์ง€ ์•Š๋Š”๋‹ค๋Š” ์„ค๋ช…์ด๋‹ค. ์• ๋ค ๋ฐฑ์ด MPC๋ฅผ ๊ฒฝ๊ณ„ํ•˜๋Š” ์ด์œ  ํฅ๋ฏธ๋กœ์šด ๋ถ€๋ถ„์€ MPC ์ œํ’ˆ์„ ๋‚ด๋†“์€ ํšŒ์‚ฌ์˜ ๋Œ€ํ‘œ๊ฐ€ MPC์˜ ๋ณต์žกํ•จ์„ ๊ฒฝ๊ณ„ํ–ˆ๋‹ค๋Š” ์ ์ด๋‹ค. ๋ฐฑ์€ ์ฝ”์ธํ…”๋ ˆ๊ทธ๋ž˜ํ”„ ์ธํ„ฐ๋ทฐ์—์„œ ์‚ฌ๋žŒ๋“ค์ด ๋งํ•˜๋Š” MPC๋Š” ์ƒ๋‹นํžˆ ๋‚œํ•ดํ•˜๊ณ  ๋ณต์žกํ•œ ํ”„๋กœํ† ์ฝœ์ด๋ผ ๋ณด์•ˆ์„ ์ง€ํ‚ค๊ธฐ ์–ด๋ ต๊ณ  ๊ธฐ๊ธฐ๋ผ๋ฆฌ ์ฃผ๊ณ ๋ฐ›๋Š” ํ†ต์‹ ๋„ ๋งŽ์•„์„œ ์™„์ „ํžˆ ์˜คํ”„๋ผ์ธ์ธ ์ฝœ๋“œ ๋ณด๊ด€์—๋Š” ์“ฐ๊ธฐ ์–ด๋ ต๋‹ค๊ณ  ๋งํ–ˆ๋‹ค. ๊ทธ๋Š” ๋ฐ”์ด๋‚ธ์Šค ํ•ดํ‚น์„ ์˜ˆ๋กœ ๋“ค์—ˆ๋‹ค. ์˜ฌํ•ด ๊ฐ€์žฅ ํฐ ์‚ฌ๊ณ ์ด๋‹ค. ๋ฐ”์ด๋‚ธ์Šค ์ž์ฒด ์กฐ์‚ฌ์— ๋”ฐ๋ฅด๋ฉด ๊ณต๊ฒฉ์ž๋Š” ์ง์› ๊ธฐ๊ธฐ๋ฅผ ์žฅ์•…ํ•ด ์–ป์€ ์ž๊ฒฉ ์ฆ๋ช…์œผ๋กœ ๊ฑฐ๋ž˜์†Œ์˜ ์ง€๊ฐ‘ ์„œ๋ช… ์ธํ”„๋ผ์— ์ ‘๊ทผํ–ˆ๋‹ค. ์ธํ„ฐ๋„ท์— ์—ฐ๊ฒฐ๋œ ์„œ๋ช… ์žฅ์น˜๋Š” ์ด๋Ÿฐ ๊ฒฝ๋กœ๋ฅผ ํ”ผํ•˜๊ธฐ ์–ด๋ ต๊ณ  ์„œ๋ช… ์žฅ์น˜๋ฅผ ์˜คํ”„๋ผ์ธ์œผ๋กœ ๋‘๋ฉด ๋ฌด๋‹จ ์ ‘๊ทผ์„ ๋ง‰๋Š” ๋ฐ ๋„์›€์ด ๋œ๋‹ค๋Š” ๊ฒƒ์ด ๊ทธ์˜ ์ฃผ์žฅ์ด๋‹ค. AI๋„ ๋ณ€์ˆ˜๋กœ ๊ผฝ์•˜๋‹ค. ๊ณต๊ฒฉ ๋„๊ตฌ๊ฐ€ ๋‹ฌ๋ผ์กŒ๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ๋ฐฑ์€ AI๊ฐ€ ๊ณต๊ฒฉ์ž์—๊ฒŒ ๋ˆˆ์— ์ž˜ ๋„์ง€ ์•Š๋Š” ์ž‘์€ ๊ตฌํ˜„ ์˜ค๋ฅ˜๋ฅผ ์ฐพ์•„ ์ฃผ๋Š” ๋„๊ตฌ๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ดค๋Š”๋ฐ ๋ณต์žกํ•œ MPC ๊ตฌํ˜„์ผ์ˆ˜๋ก ์ด๋Ÿฐ ํ‹ˆ์ด ์ƒ๊ธธ ์—ฌ์ง€๊ฐ€ ํฌ๋‹ค๋Š” ์ด์•ผ๊ธฐ์ด๋‹ค. ๊ทธ๋ž˜์„œ ๊ทธ๋Š” ๊ธฐ๊ด€๋“ค์ด ์ž์‚ฐ์„ ์ฝœ๋“œ ์Šคํ† ๋ฆฌ์ง€๋กœ ์˜ฎ๊ฒจ ํ™˜๊ฒฝ ๋ณ€ํ™”์— ๋Œ€์‘ํ•ด์•ผ ํ•œ๋‹ค๊ณ  ๊ฐ•์กฐํ–ˆ๊ณ  ๋ณต์žกํ•จ์€ ๋ณด์•ˆ์˜ ์ ์ด๋ผ๋Š” ๋ง๋กœ ์ •๋ฆฌํ–ˆ๋‹ค. ํ•ซ์›”๋ ›์€ ์ธํ„ฐ๋„ท์— ์—ฐ๊ฒฐ๋œ ์ƒํƒœ๋กœ ํ‚ค๋ฅผ ์“ฐ๋Š” ์ง€๊ฐ‘์ด๋‹ค. ์ž…์ถœ๊ธˆ์ด ๋น ๋ฅธ ๋Œ€์‹  ์›๊ฒฉ ๊ณต๊ฒฉ์— ๋…ธ์ถœ๋˜๋Š” ๋ฉด์ด ๋„“๊ณ  ๊ฑฐ๋ž˜์†Œ๊ฐ€ ์ถœ๊ธˆ ์ฒ˜๋ฆฌ์— ์“ฐ๋Š” ์ง€๊ฐ‘ ์ƒ๋‹น์ˆ˜๊ฐ€ ์—ฌ๊ธฐ์— ์†ํ•œ๋‹ค. ์ฝœ๋“œ์›”๋ ›์€ ํ‚ค๊ฐ€ ์ธํ„ฐ๋„ท๊ณผ ๋ถ„๋ฆฌ๋œ ์žฅ์น˜์— ๋จธ๋ฌด๋Š” ๋ฐฉ์‹์ด๋‹ค. ์„œ๋ช…ํ•  ๋•Œ๋„ ๊ฑฐ๋ž˜ ์ •๋ณด๋งŒ ์˜ค๊ฐ€๊ณ  ํ‚ค๋Š” ์žฅ์น˜ ๋ฐ–์œผ๋กœ ๋‚˜๊ฐ€์ง€ ์•Š๋Š”๋‹ค. ๋Œ€์‹  ์“ฐ๊ธฐ๊ฐ€ ๋ฒˆ๊ฑฐ๋กญ๊ณ  ์žฅ์น˜๋‚˜ ๋ฐฑ์—…์„ ์žƒ์œผ๋ฉด ๋˜์ฐพ๊ธฐ ์–ด๋ ต๋‹ค. ๋ฉ€ํ‹ฐ์‹œ๊ทธ๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ํ‚ค ๊ฐ€์šด๋ฐ ์ •ํ•ด์ง„ ์ˆ˜ ์ด์ƒ์ด ์„œ๋ช…ํ•ด์•ผ ์ถœ๊ธˆ๋˜๋Š” ๊ตฌ์กฐ์ธ๋ฐ ๊ทœ์น™์ด ๋ธ”๋ก์ฒด์ธ ์œ„์— ๊ทธ๋Œ€๋กœ ๋“œ๋Ÿฌ๋‚œ๋‹ค. MPC๋Š” ํ‚ค ํ•˜๋‚˜๋ฅผ ๋‚˜๋ˆ  ์„œ๋ช…์„ ๋งŒ๋“ค์–ด ๋‚ด๊ธฐ ๋•Œ๋ฌธ์— ๊ฒ‰์œผ๋กœ๋Š” ์ผ๋ฐ˜ ์„œ๋ช…์ฒ˜๋Ÿผ ๋ณด์ด๊ณ  ์šด์˜ ๋ฐฉ์‹์„ ์œ ์—ฐํ•˜๊ฒŒ ๋ฐ”๊ฟ€ ์ˆ˜ ์žˆ๋‹ค. HSM์€ ํ‚ค๋ฅผ ์ €์žฅํ•˜๊ณ  ์—ฐ์‚ฐํ•˜๋Š” ์ „์šฉ ๋ณด์•ˆ ํ•˜๋“œ์›จ์–ด๋กœ ์ด๋ฒˆ ๋ธ”๋ก์ŠคํŠธ๋ฆผ ์ œํ’ˆ์ฒ˜๋Ÿผ MPC๋‚˜ ๋‹ค๋ฅธ ๋ฐฉ์‹๊ณผ ํ•จ๊ป˜ ์“ฐ์ธ๋‹ค. ์–ด๋–ค ๋ฐฉ์‹์ด๋“  ๊ฒฐ๊ตญ ํ‚ค๊ฐ€ ์–ด๋””์—์„œ ๋งŒ๋“ค์–ด์ง€๊ณ  ์–ด๋””์—์„œ ์„œ๋ช…๋˜๋Š”์ง€๊ฐ€ ํ•ต์‹ฌ์ด๋‹ค. ๊ตฌ์กฐ๊ฐ€ ๋ณต์žกํ•ด์งˆ์ˆ˜๋ก ๊ทธ ๊ฒฝ๋กœ๋ฅผ ์ดํ•ดํ•˜๊ณ  ์ ๊ฒ€ํ•˜๋Š” ์ผ๋„ ์–ด๋ ค์›Œ์ง„๋‹ค. ํ•˜๋“œ์›จ์–ด ์ง€๊ฐ‘์„ ์“ด๋‹ค๋ฉด ์‹œ๋“œ ๋ฌธ๊ตฌ๋ฅผ ์‚ฌ์ง„์ด๋‚˜ ํด๋ผ์šฐ๋“œ ๋ฉ”๋ชจ๋กœ ๋‚จ๊ธฐ์ง€ ์•Š๋Š” ๊ฒƒ์ด ๊ธฐ๋ณธ์ด๋‹ค. ์„œ๋ช…ํ•  ๋•Œ๋Š” ์ปดํ“จํ„ฐ ํ™”๋ฉด๋ณด๋‹ค ์ง€๊ฐ‘ ์žฅ์น˜ ํ™”๋ฉด์— ๋œฌ ์ฃผ์†Œ์™€ ๊ธˆ์•ก์„ ๋จผ์ € ํ™•์ธํ•œ๋‹ค. ์ปดํ“จํ„ฐ๊ฐ€ ๊ฐ์—ผ๋˜๋”๋ผ๋„ ์žฅ์น˜ ํ™”๋ฉด์€ ๋”ฐ๋กœ ํ™•์ธํ•  ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. ์ž์ฃผ ์“ฐ๋Š” ์†Œ์•ก๊ณผ ์˜ค๋ž˜ ๋“ค๊ณ  ๊ฐˆ ์ž์‚ฐ์„ ๋‚˜๋ˆ„๋Š” ๊ฒƒ๋„ ๋ฐฉ๋ฒ•์ด๋‹ค. ๊ฑฐ๋ž˜๋‚˜ ๋””ํŒŒ์ด์— ์“ฐ๋Š” ์ง€๊ฐ‘๊ณผ ์žฅ๊ธฐ ๋ณด๊ด€ ์ง€๊ฐ‘์„ ๋ถ„๋ฆฌํ•˜๋ฉด ํ•œ์ชฝ์ด ํ„ธ๋ฆฌ๋”๋ผ๋„ ํ”ผํ•ด๋ฅผ ์ค„์ผ ์ˆ˜ ์žˆ๋‹ค. ์ด๋ฒˆ ๋ฐœํ‘œ๋Š” ๊ธฐ๊ด€ ์‹œ์žฅ์„ ๊ฒจ๋ƒฅํ•œ ์ œํ’ˆ์ด์ง€๋งŒ ๊ทธ ๋ฐ”ํƒ•์— ๊น”๋ฆฐ ๋ฉ”์‹œ์ง€๋Š” ๋‹จ์ˆœํ•˜๋‹ค. ํ‚ค๋ฅผ ์ธํ„ฐ๋„ท์—์„œ ๋ฉ€๋ฆฌ ๋‘๊ณ  ๊ตฌ์กฐ๋Š” ์ดํ•ดํ•  ์ˆ˜ ์žˆ์„ ๋งŒํผ๋งŒ ๋ณต์žกํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด ์ง€๊ธˆ์ฒ˜๋Ÿผ ๊ณต๊ฒฉ์ด ์ •๊ตํ•ด์ง„ ์‹œ๊ธฐ์— ๊ฐ€์žฅ ๊ธฐ๋ณธ์ ์ธ ๋Œ€์‘์ด๋ผ๋Š” ์ด์•ผ๊ธฐ์ด๋‹ค.
66
๋ธ”๋ฃธ๋ฒ„๊ทธ๊ฐ€ ํ˜„์ง€์‹œ๊ฐ 13์ผ ์ฝ”์ธ๊ธ€๋ž˜์Šค์˜ ์„ ๋ฌผ๊ฑฐ๋ž˜์†Œ ์Šค๋ฌผ์—ฌ๋Ÿ ๊ณณ ์ž๋ฃŒ๋ฅผ ์ธ์šฉํ•œ ๋ณด๋„์— ๋”ฐ๋ฅด๋ฉด ์ตœ๊ทผ ์ผ ๋…„ ๋™์•ˆ ๋น„ํŠธ์ฝ”์ธ ํฌ์ง€์…˜ ์ฒญ์‚ฐ์•ก์ด ํ•˜๋ฃจ ์˜ค์–ต ๋‹ฌ๋Ÿฌ๋ฅผ ๋„˜๊ธด ๋‚ ์ด ์—ด์‚ฌํ˜์ด๋‚˜ ๋๊ณ  ๊ทธ ๊ฐ€์šด๋ฐ ์‚ฌํ˜์€ ํ•˜๋ฃจ ์ฒญ์‚ฐ์ด ์‹ญ์–ต ๋‹ฌ๋Ÿฌ๋ฅผ ๋„˜์—ˆ๋‹ค. ์ด๋ฒˆ ์ฃผ๋„ ์˜ˆ์™ธ๋Š” ์•„๋‹ˆ์—ˆ๋Š”๋ฐ ๋ธ”๋ก๋ฏธ๋””์–ด๊ฐ€ ์ „ํ•œ ๋ธ”๋ฃธ๋ฒ„๊ทธ ๋ณด๋„์— ๋”ฐ๋ฅด๋ฉด ์ง€๋‚œ ํ•˜๋ฃจ ๋™์•ˆ ์ฒญ์‚ฐ๋œ ์ƒ์Šน ๋ฒ ํŒ…์€ ์•ฝ ์‹ญ์‚ผ์–ต ๋‹ฌ๋Ÿฌ์˜€๊ณ  ๋Œ€๋ถ€๋ถ„์ด ๋น„ํŠธ์ฝ”์ธ ํฌ์ง€์…˜์ด์—ˆ๋‹ค. ํŒŒ์ƒ์ƒํ’ˆ์€ ์ง€๊ธˆ ๋””์ง€ํ„ธ์ž์‚ฐ ์‹œ์žฅ ์ „์ฒด ๊ฑฐ๋ž˜๋Ÿ‰์˜ ์‚ฌ๋ถ„์˜ ์‚ผ ์ •๋„๋ฅผ ์ฐจ์ง€ํ•œ๋‹ค. ์œˆ์„ผํŠธ์˜ ํด ํ•˜์›Œ๋“œ๋Š” ์ตœ๊ทผ ์—ด๋‘ ๋‹ฌ ์‚ฌ์ด ๋Œ€๊ทœ๋ชจ ์ฒญ์‚ฐ์ด ์ „๋ณด๋‹ค ์žฆ์•„์กŒ๋‹ค๊ณ  ์„ค๋ช…ํ–ˆ๋‹ค. ๋†’์€ ์ˆ˜์ต์„ ๋…ธ๋ฆฌ๊ณ  ๋ ˆ๋ฒ„๋ฆฌ์ง€๋ฅผ ํ‚ค์šด ์ƒํƒœ์—์„œ ๊ฑฐ์‹œ๊ฒฝ์ œ ๋‰ด์Šค๊ฐ€ ๋‚˜์˜ค๋ฉด ํ•œ๊บผ๋ฒˆ์— ์ฒญ์‚ฐ์ด ํ„ฐ์ง„๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ๊ฒฐ๊ตญ ๋ ˆ๋ฒ„๋ฆฌ์ง€ ์ด์•ผ๊ธฐ์ด๋‹ค. ์‹œ๊ทธ๋„˜์€ํ–‰์˜ ํŒŒ๋น„์•ˆ ๋„๋ฆฌ๋Š” ๋‹ค๋ฅธ ์‹œ๊ฐ์„ ๋‚ด๋†จ๋Š”๋ฐ ํฐ ์ฒญ์‚ฐ์ด ๊ณง ๋น„ํŠธ์ฝ”์ธ ์ž์ฒด์˜ ์•ฝ์ ์„ ๋œปํ•˜์ง€๋Š” ์•Š๋Š”๋‹ค๋Š” ๊ฒƒ์ด๋‹ค. ์ฒญ์‚ฐ ๊ทœ๋ชจ๋Š” ์‹œ์žฅ์— ๋ ˆ๋ฒ„๋ฆฌ์ง€๊ฐ€ ์–ด๋–ป๊ฒŒ ๊น”๋ ค ์žˆ๋А๋ƒ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ์„ค๋ช…์ด๋‹ค. ๋…ธ์—˜ ์• ์น˜์Šจ์€ ETF ์ดํ›„ ๋น„ํŠธ์ฝ”์ธ์ด ๊ฑฐ์‹œ ์ž์‚ฐ์ด ๋˜๋ฉด์„œ ๋‹ค๋ฅธ ์ž์‚ฐ๊ณผ ์ž๊ธˆ์„ ๋‘๊ณ  ๊ฒฝ์Ÿํ•˜๊ฒŒ ๋๋‹ค๊ณ  ์งš์—ˆ๋‹ค. ํ•ด์„์€ ์กฐ๊ธˆ์”ฉ ๊ฐˆ๋ฆฐ๋‹ค. ํ•˜๋ฃจ ๋ช‡์–ต ๋‹ฌ๋Ÿฌ์”ฉ ์ฒญ์‚ฐ์ด ์Ÿ์•„์ง€๋Š” ๋‚ ์€ ๊ฐ€๊ฒฉ์ด ๊ฐ€๋Š” ๋ฐฉํ–ฅ๋ณด๋‹ค ๋ ˆ๋ฒ„๋ฆฌ์ง€๊ฐ€ ์–ด๋””์— ๋ชฐ๋ ค ์žˆ์—ˆ๋Š”์ง€๋ฅผ ๋จผ์ € ๋ณด์—ฌ ์ค€๋‹ค. ์„ ๋ฌผ์„ ์“ด๋‹ค๋ฉด ์ง€๊ธˆ์ฒ˜๋Ÿผ ๊ฑฐ์‹œ ์ง€ํ‘œ์™€ ๋‰ด์Šค๊ฐ€ ๋ชฐ๋ฆฌ๋Š” ์ฃผ์—๋Š” ๋ฐฐ์œจ๊ณผ ์ฒญ์‚ฐ ๊ฐ€๊ฒฉ๋ถ€ํ„ฐ ๋‹ค์‹œ ํ™•์ธํ•˜๋Š” ํŽธ์ด ๋‚ซ๊ณ  ํ˜„๋ฌผ๋งŒ ๋“ค๊ณ  ์žˆ๋”๋ผ๋„ ์ด๋Ÿฐ ๋‚ ์˜ ๊ธ‰๋ฝ์ด ๋ ˆ๋ฒ„๋ฆฌ์ง€ ์ •๋ฆฌ์—์„œ ๋‚˜์˜จ ๊ฒƒ์ธ์ง€ ๋”ฐ๋กœ ๊ตฌ๋ถ„ํ•ด์„œ ๋ณด๋Š” ์Šต๊ด€์ด ๋„์›€์ด ๋œ๋‹ค.
1
1
65
๐Ÿ“š Rialo Research Series 62: Financial Remediation - Can Autonomous Finance Repair the Damage? (Kor : ๊ธˆ์œต ํ”ผํ•ด ๋ณต๊ตฌ - ์ž์œจ ๊ธˆ์œต ์‹œ์Šคํ…œ์€ ์ž˜๋ชป๋œ ๊ฑฐ๋ž˜์˜ ํ”ผํ•ด๊นŒ์ง€ ๋ณต๊ตฌํ•  ์ˆ˜ ์žˆ์„๊นŒ?) 1/ Intro: Correcting Trust Doesn't Reverse Losses (EN) In the previous chapter, we explored how an agent might recover from manipulated confidence scores. But restoring trust doesn't undo financial damage. Imagine a lending protocol liquidating a healthy position because corrupted oracle data made its collateral appear insufficient. The oracle is later corrected, but the collateral has already been sold. Fixing the decision model and repairing its consequences are separate problems. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ์กฐ์ž‘๋œ Confidence Score๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์‹ ๋ขฐ๋„๋ฅผ ๋ฐ”๋กœ์žก๋Š”๋‹ค๊ณ  ์ด๋ฏธ ๋ฐœ์ƒํ•œ ๊ธˆ์œต ์†์‹ค๊นŒ์ง€ ์‚ฌ๋ผ์ง€๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ž˜๋ชป๋œ Oracle ๊ฐ€๊ฒฉ ๋•Œ๋ฌธ์— ์ •์ƒ์ ์ธ ๋‹ด๋ณด ํฌ์ง€์…˜์ด Liquidation๋๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ดํ›„ ๊ฐ€๊ฒฉ ์ •๋ณด๊ฐ€ ์ˆ˜์ •๋˜๋”๋ผ๋„ ๋‹ด๋ณด๋Š” ์ด๋ฏธ ๋งค๊ฐ๋์Šต๋‹ˆ๋‹ค. ํŒ๋‹จ ๊ธฐ์ค€์„ ๋ณต๊ตฌํ•˜๋Š” ๊ฒƒ๊ณผ ์‹ค์ œ ํ”ผํ•ด๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. 2/ Why Transactions Cannot Simply Be Reversed (EN) Once a blockchain transaction is finalized, an application generally cannot erase it as though it never happened. A liquidation may involve collateral transfers, debt repayment, and third-party buyers. Reversing one operation could require undoing several dependent transactions. Financial remediation therefore often means new compensating actions rather than rewriting history. (KR) Blockchain Transaction์ด Finalized๋˜๋ฉด ์ผ๋ฐ˜์ ์œผ๋กœ ์‹คํ–‰ ๊ธฐ๋ก์„ ์—†์—ˆ๋˜ ์ผ์ฒ˜๋Ÿผ ๋˜๋Œ๋ฆด ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค. Liquidation ๊ณผ์ •์—๋Š” ๋‹ด๋ณด ์ด์ „, ๋ถ€์ฑ„ ์ƒํ™˜, ์ œ3์ž์˜ ์ž์‚ฐ ๋งค์ž… ๋“ฑ์ด ํฌํ•จ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜๋‚˜์˜ ๊ฑฐ๋ž˜๋ฅผ ๋˜๋Œ๋ฆฌ๋ ค๋ฉด ์—ฐ๊ฒฐ๋œ ๋‹ค๋ฅธ ๊ฑฐ๋ž˜๊นŒ์ง€ ์˜ํ–ฅ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ”ผํ•ด ๋ณต๊ตฌ๋Š” ๊ณผ๊ฑฐ ๊ธฐ๋ก์„ ์‚ญ์ œํ•˜๊ธฐ๋ณด๋‹ค ์ƒˆ๋กœ์šด ๋ณด์ƒ Transaction์„ ์‹คํ–‰ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ ‘๊ทผํ•ด์•ผ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 3/ Compensating Transactions (EN) Suppose a user loses $5,000 through an incorrect liquidation. A protocol could compensate the user from a designated reserve, subject to predefined eligibility and approval rules. But compensation is not identical to reversal. Market prices may have changed, and the original assets may no longer be available. The policy must define what is compensated and how losses are calculated. (KR) ์ž˜๋ชป๋œ Liquidation์œผ๋กœ ์‚ฌ์šฉ์ž๊ฐ€ 5,000๋‹ฌ๋Ÿฌ์˜ ์†์‹ค์„ ์ž…์—ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. Protocol์€ ์‚ฌ์ „์— ๋งˆ๋ จํ•œ Reserve๋ฅผ ํ™œ์šฉํ•ด ๋ณด์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋งŒ ๋ณด์ƒ ๋Œ€์ƒ๊ณผ ์Šน์ธ ์กฐ๊ฑด์€ ๋ฏธ๋ฆฌ ์ •์˜๋ผ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ณด์ƒ์€ ์›์ƒ๋ณต๊ตฌ์™€ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ์ž์‚ฐ ๊ฐ€๊ฒฉ์ด ๋ณ€ํ–ˆ๊ฑฐ๋‚˜ ๊ธฐ์กด ๋‹ด๋ณด๋ฅผ ๋‹ค์‹œ ํ™•๋ณดํ•  ์ˆ˜ ์—†์„ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ์–ด๋–ค ์†์‹ค์„ ์–ด๋–ค ๊ธฐ์ค€์œผ๋กœ ๋ณด์ƒํ• ์ง€ ๋ช…ํ™•ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 4/ The Risk of False Compensation Claims (EN) Remediation mechanisms can become attack targets. An attacker might deliberately create suspicious transactions and claim they were harmed by an oracle failure. If compensation relies only on an agent's explanation, fraudulent claims may be difficult to distinguish from genuine losses. Eligibility should be checked against historical state, execution records, and incident-specific evidence. (KR) ํ”ผํ•ด ๋ณด์ƒ ๊ธฐ๋Šฅ ์ž์ฒด๊ฐ€ ๊ณต๊ฒฉ ๋Œ€์ƒ์ด ๋  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์ž๊ฐ€ ์˜๋„์ ์œผ๋กœ ๋น„์ •์ƒ์ ์ธ ๊ฑฐ๋ž˜๋ฅผ ๋งŒ๋“  ๋’ค Oracle ์žฅ์• ๋กœ ์†ํ•ด๋ฅผ ์ž…์—ˆ๋‹ค๊ณ  ์ฃผ์žฅํ•˜๋Š” ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. Agent์˜ ์„ค๋ช…๋งŒ์œผ๋กœ ๋ณด์ƒํ•œ๋‹ค๋ฉด ํ—ˆ์œ„ ์ฒญ๊ตฌ๋ฅผ ๊ตฌ๋ถ„ํ•˜๊ธฐ ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Historical State, Transaction Record, ์žฅ์•  ๋‹น์‹œ์˜ Evidence๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๋ณด์ƒ ์ž๊ฒฉ์„ ๊ฒ€์ฆํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 5/ Who Pays for the Damage? (EN) Even when an incorrect liquidation is proven, responsibility may remain disputed. Was the loss caused by a faulty oracle, an application bug, or an agent using outdated data? Possible arrangements include insurance reserves, protocol-funded compensation, or contractual liability. None automatically follows from the execution record. Responsibility must be established through governance, agreements, and applicable rules. (KR) ์ž˜๋ชป๋œ Liquidation์ด ํ™•์ธ๋ผ๋„ ๋ˆ„๊ฐ€ ์†์‹ค์„ ๋ถ€๋‹ดํ• ์ง€๋Š” ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. Oracle ์˜ค๋ฅ˜์ธ์ง€, Smart Contract Bug์ธ์ง€, Agent๊ฐ€ ์˜ค๋ž˜๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ ๊ฒƒ์ธ์ง€์— ๋”ฐ๋ผ ์ฑ…์ž„ ๊ตฌ์กฐ๊ฐ€ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Insurance Reserve๋‚˜ Protocol Fund, ์‚ฌ์ „์— ์ •์˜๋œ ๊ณ„์•ฝ์ƒ ์ฑ…์ž„ ๋“ฑ์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹คํ–‰ ๊ธฐ๋ก์ด ์›์ธ์„ ๋ณด์—ฌ์ฃผ๋”๋ผ๋„ ๋ณด์ƒ ์ฑ…์ž„๊นŒ์ง€ ์ž๋™์œผ๋กœ ๊ฒฐ์ •ํ•ด ์ฃผ์ง€๋Š” ์•Š์Šต๋‹ˆ๋‹ค. 6/ Rialo and Automated Remediation (EN) Rialo's Reactive Transactions model supports condition-based transaction triggering during block execution. Applications could use reactive workflows to detect qualifying incidents and initiate predefined remediation procedures. However, reactive execution does not automatically reverse finalized transactions or determine compensation liability. Loss calculations, approvals, funding, and dispute resolution remain application-level responsibilities. (KR) Rialo์˜ Reactive Transactions๋ฅผ ํ™œ์šฉํ•˜๋ฉด ํŠน์ • ์žฅ์•  ์กฐ๊ฑด์ด ํ™•์ธ๋์„ ๋•Œ ์‚ฌ์ „์— ์ •์˜๋œ Remediation Workflow๋ฅผ ์‹œ์ž‘ํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ์„ค๊ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Reactive Execution์ด Finalized Transaction์„ ์ž๋™์œผ๋กœ ์ทจ์†Œํ•˜๊ฑฐ๋‚˜ ๋ณด์ƒ ์ฑ…์ž„์„ ๊ฒฐ์ •ํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์†์‹ค ๊ณ„์‚ฐ, ๋ณด์ƒ ์Šน์ธ, ์žฌ์› ํ™•๋ณด, ๋ถ„์Ÿ ํ•ด๊ฒฐ์€ Application์—์„œ ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ž๋™ ์‹คํ–‰๊ณผ ์ž๋™ ๋ณด์ƒ์€ ๊ฐ™์€ ๊ฐœ๋…์ด ์•„๋‹™๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a reactive lending demonstration where manipulated oracle data causes an incorrect liquidation. After detecting the error, the application should reconstruct the historical state, calculate the verified loss, and process a controlled compensation request. The demonstration should also reject a fraudulent claim based on unrelated transactions. This would show how verifiable evidence can support remediation without promising impossible transaction reversals. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ Demo๋Š” ์กฐ์ž‘๋œ Oracle ๋ฐ์ดํ„ฐ๋กœ ์ž˜๋ชป๋œ Liquidation์ด ๋ฐœ์ƒํ•˜๋Š” ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. ์˜ค๋ฅ˜ ๋ฐœ๊ฒฌ ํ›„ Historical State๋ฅผ ์žฌ๊ตฌ์„ฑํ•˜๊ณ  ์‹ค์ œ ์†์‹ค์„ ๊ณ„์‚ฐํ•œ ๋’ค ๊ฒ€์ฆ๋œ ๋ณด์ƒ ์š”์ฒญ์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ณผ์ •์„ ํ™•์ธํ•˜๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ๋™์‹œ์— ๋ฌด๊ด€ํ•œ ๊ฑฐ๋ž˜๋ฅผ ์ด์šฉํ•œ ํ—ˆ์œ„ ๋ณด์ƒ ์ฒญ๊ตฌ๋Š” ๊ฑฐ๋ถ€๋ผ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํ”ผํ•ด ๋ณต๊ตฌ์˜ ํ•ต์‹ฌ์€ ๊ฑฐ๋ž˜๋ฅผ ๋˜๋Œ๋ฆฌ๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ๊ทผ๊ฑฐ๋กœ ์†์‹ค์„ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. (EN) We've explored why financial remediation requires evidence, compensation rules, and clear responsibility. But if several agents contribute to one incorrect decision, how should an autonomous system divide accountability among them? (KR) ์ด๋ฒˆ ํŽธ์—์„œ๋Š” Financial Remediation์„ ์ค‘์‹ฌ์œผ๋กœ ์ž˜๋ชป๋œ ์ž๋™ํ™” ํŒ๋‹จ์ด ์‹ค์ œ ๊ธˆ์œต ์†์‹ค๋กœ ์ด์–ด์กŒ์„ ๋•Œ ํ•„์š”ํ•œ ๋ณด์ƒ๊ณผ ๊ฒ€์ฆ ์ ˆ์ฐจ๋ฅผ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์—ฌ๋Ÿฌ Agent๊ฐ€ ํ•˜๋‚˜์˜ ์ž˜๋ชป๋œ ๊ฒฐ์ •์— ๊ด€์—ฌํ–ˆ๋‹ค๋ฉด ์ฑ…์ž„์€ ์–ด๋–ป๊ฒŒ ๋‚˜๋ˆ ์•ผ ํ• ๊นŒ์š”? Oracle ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•œ Agent, ์œ„ํ—˜์„ ํ‰๊ฐ€ํ•œ Agent, Transaction์„ ์‹คํ–‰ํ•œ Agent๊ฐ€ ๋ชจ๋‘ ๋‹ค๋ฅธ ์—ญํ• ์„ ์ˆ˜ํ–‰ํ–ˆ์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Autonomous Finance์—์„œ๋Š” ์‹คํ–‰ ๊ณผ์ •์˜ ์—ญํ• ๋ฟ ์•„๋‹ˆ๋ผ ์‹คํŒจ์— ๋Œ€ํ•œ ์ฑ…์ž„๊นŒ์ง€ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•˜์ง€ ์•Š์„๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
1
140
ํ•˜๋ฝ์žฅ์—์„œ ๋“œ๋Ÿฌ๋‚œ ํŠธ๋ ˆ์ €๋ฆฌ ๊ธฐ์—…์˜ ๋‹ด๋ณด ๋Œ€์ถœ, BTCS์™€ ๋‚˜์นด๋ชจํ†  ๊ณต์‹œ๋กœ ๋ณธ ๊ตฌ์กฐ ์ฝ”์ธ์„ ํšŒ์‚ฌ ์ž์‚ฐ์œผ๋กœ ์Œ“์•„ ์˜จ ์ƒ์žฅ์‚ฌ ๋‘ ๊ณณ์ด ์–ด์ œ ๋ฏธ๊ตญ์—์„œ ๊ณต์‹œ๋ฅผ ๋ƒˆ๋Š”๋ฐ ๋‘˜ ๋‹ค ๊ฐ€๊ฒฉ์ด ๋‚ด๋ ค๊ฐ€๋Š” ๊ตญ๋ฉด์—์„œ ๋‹ด๋ณด ๋Œ€์ถœ์„ ์–ด๋–ป๊ฒŒ ๋ฒ„ํ‹ฐ๊ณ  ์žˆ๋Š”์ง€๋ฅผ ๋ณด์—ฌ ์ค€๋‹ค. ์ด๋”๋ฆฌ์›€์„ ์ฃผ๋กœ ๋“ค๊ณ  ์žˆ๋Š” BTCS๋Š” ์—์ด๋ธŒ ๋Œ€์ถœ์„ ์ง€ํ‚ค๋ ค๊ณ  ์‚ฐ ๊ฐ’๋ณด๋‹ค ์‹ธ๊ฒŒ ์ด๋”๋ฆฌ์›€์„ ํŒ”์•˜๊ณ  ๋น„ํŠธ์ฝ”์ธ์„ ๋“ค๊ณ  ์žˆ๋Š” ๋‚˜์นด๋ชจํ† ๋Š” ํฌ๋ผ์ผ„์— ์ง„ ๋นš์˜ ๋งŒ๊ธฐ๋ฅผ ๋‚ด๋…„์œผ๋กœ ๋ฏธ๋ค˜๋‹ค. ๋‘ ํšŒ์‚ฌ ๋ชจ๋‘ ์•„์ง ์ฒญ์‚ฐ์„ ๋‹นํ•˜์ง€๋Š” ์•Š์•˜๋‹ค. ๋‹ค๋งŒ ๊ณต์‹œ ์ˆซ์ž๋ฅผ ๋”ฐ๋ผ๊ฐ€ ๋ณด๋ฉด ๋‹ด๋ณด ๋Œ€์ถœ๋กœ ์ฝ”์ธ์„ ๋” ์‚ฌ ์˜จ ๊ตฌ์กฐ๊ฐ€ ํ•˜๋ฝ์žฅ์—์„œ ์–ด๋””๊ฐ€ ์•ฝํ•œ์ง€ ๊ฝค ์„ ๋ช…ํ•˜๊ฒŒ ๋“œ๋Ÿฌ๋‚œ๋‹ค. BTCS๋Š” ๋ถ„๊ธฐ ๋ณด๊ณ ์„œ์—์„œ ์ด๋‹ฌ ์ดˆ ๊ธฐ์ค€ ์—์ด๋ธŒ์— ์ด์ž๋ฅผ ํฌํ•จํ•ด ์•ฝ ์œก์ฒœ์‚ฌ๋ฐฑ๋งŒ ๋‹ฌ๋Ÿฌ๋ฅผ ๋นŒ๋ ค ์“ฐ๊ณ  ์žˆ๋‹ค๊ณ  ๋ฐํ˜”๋‹ค. ๋‹ด๋ณด๋Š” ์ด๋”๋ฆฌ์›€ ์‚ฌ๋งŒ ์œก์ฒœ์—ฌ ๊ฐœ์˜€๋‹ค. ํšŒ์‚ฌ๊ฐ€ ์ ์šฉํ•œ ๊ฐ€๊ฒฉ์œผ๋กœ ๊ณ„์‚ฐํ•˜๋ฉด ์•ฝ ๊ตฌ์ฒœ๊ตฌ๋ฐฑ๋งŒ ๋‹ฌ๋Ÿฌ๋กœ ๋นš์˜ ํ•œ ๋ฐฐ ๋ฐ˜ ๋‚จ์ง“์ด๋‹ค. ์—์ด๋ธŒ๋Š” ๋Œ€์ถœ ๋‹ด๋‹น์ž๊ฐ€ ๋”ฐ๋กœ ์—†๋Š” ํ”„๋กœํ† ์ฝœ์ด๋‹ค. ๋Œ€์‹  ๋‹ด๋ณด ๊ฐ€์น˜์™€ ๋นš์„ ๋น„๊ตํ•œ ๊ฑด์ „์„ฑ ์ง€ํ‘œ๋ฅผ ๊ณ„์† ๊ณ„์‚ฐํ•˜๊ณ  ์ด ๊ฐ’์ด ์ผ ๋ฐ‘์œผ๋กœ ๋–จ์–ด์ง€๋ฉด ๋‹ด๋ณด ์ผ๋ถ€๋ฅผ ์ž๋™์œผ๋กœ ํŒ”์•„ ๋นš์„ ๊ฐš๋Š”๋‹ค. ์ฝ”์ธ๋ฐ์Šคํฌ๋Š” ์—์ด๋ธŒ๊ฐ€ BTCS์˜ ์ด๋”๋ฆฌ์›€ ๊ฐ€์น˜๋ฅผ ์•ฝ 8ํ• ๋งŒ ์ธ์ •ํ•œ๋‹ค๋Š” ํšŒ์‚ฌ ์„ค๋ช…์„ ๋ฐ”ํƒ•์œผ๋กœ ์ฒญ์‚ฐ์„ ์ด ์ด๋”๋ฆฌ์›€ 1,717๋‹ฌ๋Ÿฌ ๋ถ€๊ทผ์ด๋ผ๊ณ  ๊ณ„์‚ฐํ–ˆ๋Š”๋ฐ ์˜ค๋Š˜ ์ด๋”๋ฆฌ์›€ ๊ฐ€๊ฒฉ์€ ๊ทธ๋ณด๋‹ค 1ํ•  ๋‚จ์ง“ ์œ„์— ์žˆ๋‹ค. ํšŒ์‚ฌ๋Š” ์ด๋”๋ฆฌ์›€ ๋งค๋„๊ฐ€ ๋‹ด๋ณด ์ˆ˜์ค€๊ณผ ๊ฑด์ „์„ฑ ์ง€ํ‘œ๋ฅผ ๊ด€๋ฆฌํ•˜๋Š” ๊ณผ์ •์—์„œ ์›๊ฐ€๋ณด๋‹ค ๋‚ฎ์€ ๊ฐ€๊ฒฉ์— ์ด๋ค„์กŒ๋‹ค๊ณ  ์ ์—ˆ๋‹ค. ์–ผ๋งˆ๋ฅผ ํŒ”์•˜๋Š”์ง€๋Š” ๋ฐํžˆ์ง€ ์•Š์•˜๋‹ค. 9์›” ๋ง ํ˜„๊ธˆ์€ ์˜ค์‹ญ๋งŒ ๋‹ฌ๋Ÿฌ๊ฐ€ ์ฑ„ ์•ˆ ๋๊ณ  ๋””์ง€ํ„ธ ์ž์‚ฐ์˜ ์ƒ๋‹น ๋ถ€๋ถ„์ด ์—์ด๋ธŒ์— ๋‹ด๋ณด๋กœ ๋ฌถ์—ฌ ์žˆ์–ด ์“ธ ์ˆ˜ ์žˆ๋Š” ์—ฌ์œ ๊ฐ€ ํฌ์ง€ ์•Š๋‹ค. ๋นš์ด ๋Š˜์–ด๋‚œ ์†๋„๋„ ๋ˆˆ์— ๋ˆ๋‹ค. ์ผ ๋…„ ๋‚จ์ง“ ์‚ฌ์ด์˜ ์ผ์ด๋‹ค. ์ฒ˜์Œ ์—์ด๋ธŒ ๋Œ€์ถœ์„ ๊ณต๊ฐœํ•œ ์ง€๋‚œํ•ด 4์›”์—๋Š” ๋นŒ๋ฆฐ ๋ˆ์ด ์ˆ˜์‹ญ๋งŒ ๋‹ฌ๋Ÿฌ ์ˆ˜์ค€์ด์—ˆ๋Š”๋ฐ ์ง€๋‚œ 8์›”์—๋Š” ์˜ค์ฒœ๋งŒ ๋‹ฌ๋Ÿฌ๋ฅผ ๋„˜๊ฒผ๋‹ค. ์ด์‚ฌํšŒ๋Š” ๋‹ด๋ณด ๊ฐ€์น˜์˜ ์ ˆ๋ฐ˜๊นŒ์ง€๋งŒ ๋นŒ๋ฆฌ๋„๋ก ํ•œ๋„๋ฅผ ์ •ํ•ด ๋’€์ง€๋งŒ ์ด๋”๋ฆฌ์›€ ๊ฐ€๊ฒฉ์ด ๋‚ด๋ ค๊ฐ€๋ฉด ์ด ๋น„์œจ์€ ์ €์ ˆ๋กœ ์˜ฌ๋ผ๊ฐ€๊ณ  ๊ทธ๋•Œ๋งˆ๋‹ค ๋‹ด๋ณด๋ฅผ ๋” ๋„ฃ๊ฑฐ๋‚˜ ๋นš์„ ๊ฐš๊ฑฐ๋‚˜ ์ž์‚ฐ์„ ํŒ”์•„์•ผ ํ•œ๋‹ค. ๋‚˜์นด๋ชจํ† ๊ฐ€ ๋งŒ๊ธฐ๋ฅผ ๋Šฆ์ถฐ ๋ฒˆ ์‹œ๊ฐ„ ๋‚˜์นด๋ชจํ† ๋Š” ํฌ๋ผ์ผ„์— ์ผ์–ต ์œก์ฒœ์˜ค๋ฐฑ๋งŒ USDT๋ฅผ ๋นš์ง€๊ณ  ์žˆ๋‹ค. ์ด ๋Œ€์ถœ์€ ๋น„ํŠธ์ฝ”์ธ๋งŒ์„ ๋‹ด๋ณด๋กœ ์žก๊ณ  ์žˆ๋Š”๋ฐ ์–ด์ œ ์ข…๊ฐ€ ๊ธฐ์ค€ ํšŒ์‚ฌ ์‹œ๊ฐ€์ด์•ก์€ ์•ฝ ์ผ์–ต ์‚ฌ๋ฐฑ๋งŒ ๋‹ฌ๋Ÿฌ๋กœ ๋นš๋ณด๋‹ค ์ž‘๋‹ค. ์ด๋ฒˆ์— ๊ณต์‹œ๋œ ๋ณ€๊ฒฝ ๊ณ„์•ฝ์œผ๋กœ ์›๋ž˜ ๋‹ค์Œ ๋‹ฌ ์ดˆ์˜€๋˜ ์œก์ฒœ๋งŒ USDT์˜ ๋งŒ๊ธฐ๊ฐ€ ๋‚ด๋…„ 12์›”๋กœ ๋ฐ€๋ ธ๋‹ค. ๋‚˜๋จธ์ง€ ๋นš์˜ ๋งŒ๊ธฐ๋Š” ๋‚ด๋…„ 6์›” ๋ง ๊ทธ๋Œ€๋กœ๋ผ์„œ ์˜ฌํ•ด ์•ˆ์— ๊ฐš์„ ๋ˆ์€ ์—†์–ด์กŒ๋‹ค. ๋‹ด๋ณด ์—ฌ์œ ๋Š” ๋„‰๋„‰ํ•˜์ง€ ์•Š๋‹ค. 9์›” ๋ง ๊ธฐ์ค€ ๋ณด์œ  ๋น„ํŠธ์ฝ”์ธ์˜ ์น  ํ•  ์ •๋„๊ฐ€ ํฌ๋ผ์ผ„์— ๋‹ด๋ณด๋กœ ์žกํ˜€ ์žˆ์—ˆ๋Š”๋ฐ ์ฝ”์ธ๋ฐ์Šคํฌ ๊ณ„์‚ฐ์œผ๋กœ๋Š” ์ง€๊ธˆ ๊ฐ€๊ฒฉ์—์„œ ์ด ๋‹ด๋ณด ๊ฐ€์น˜๊ฐ€ ๋นš๋ณด๋‹ค ์‚ฌ๋ถ„์˜ ์ผ ์ •๋„ ๋งŽ์€ ์ˆ˜์ค€์— ๊ทธ์นœ๋‹ค. ๋‹ด๋ณด ๊ฐ€์น˜๊ฐ€ ์ •ํ•ด์ง„ ์œ ์ง€ ๊ธฐ์ค€ ์•„๋ž˜๋กœ ๋–จ์–ด์ง€๋ฉด ๋นš ์ผ๋ถ€๋ฅผ ๊ฐš๊ฑฐ๋‚˜ ๋น„ํŠธ์ฝ”์ธ์„ ๋” ๋งก๊ฒจ์•ผ ํ•˜๊ณ  ๋‹ด๋ณด๋กœ ์žกํžˆ์ง€ ์•Š์€ ๋น„ํŠธ์ฝ”์ธ์€ ์ฒœ์ด๋ฐฑ์—ฌ ๊ฐœ๊ฐ€ ๋‚จ์•„ ์žˆ๋‹ค. ํฌ๋ผ์ผ„์€ ์ด๋ฒˆ ๋ณ€๊ฒฝ์—์„œ ๋‚˜์นด๋ชจํ† ์˜ ํŒŒ์ƒ์ƒํ’ˆ ๊ณ„์ขŒ ์ˆœ๊ฐ€์น˜๋„ ๋‹ด๋ณด ํ‰๊ฐ€์— ๋„ฃ์–ด ์ฃผ๊ธฐ๋กœ ํ–ˆ๋‹ค. ๋‚˜์นด๋ชจํ† ๋Š” ๊ฐ€๊ฒฉ ํ•˜๋ฝ์— ๋Œ€๋น„ํ•œ ํ’‹์˜ต์…˜์„ ๋“ค๊ณ  ์žˆ๋Š”๋ฐ ์ด ์˜ต์…˜์€ 12์›” 25์ผ์— ๋งŒ๊ธฐ๊ฐ€ ๋Œ์•„์™€ ๋Œ€์ถœ ๋งŒ๊ธฐ๋ณด๋‹ค ๋จผ์ € ๋๋‚œ๋‹ค. ๊ณ„์ขŒ ์ˆœ๊ฐ€์น˜๊ฐ€ ๋งˆ์ด๋„ˆ์Šค์ด๋ฉด ์˜์œผ๋กœ ๊ณ„์‚ฐ๋œ๋‹ค๋Š” ์กฐ๊ฑด๋„ ๋ถ™์—ˆ๋‹ค. ๋‘ ํšŒ์‚ฌ ๋ชจ๋‘ ์ฝ”์ธ ๊ฐ€๊ฒฉ์ด ์˜ค๋ฅผ ๋•Œ ๋นŒ๋ฆฐ ๋ˆ์œผ๋กœ ์ฝ”์ธ์„ ๋” ์‚ฌ๋“ค์ธ ๊ตฌ์กฐ์ด๋‹ค. ๊ฐ€๊ฒฉ์ด ์˜ค๋ฅด๋Š” ๋™์•ˆ์—๋Š” ๋‹ด๋ณด ๊ฐ€์น˜๋„ ํ•จ๊ป˜ ์˜ฌ๋ผ ๋ฌธ์ œ๊ฐ€ ๋ณด์ด์ง€ ์•Š๋Š”๋ฐ ๋ฐฉํ–ฅ์ด ๋ฐ”๋€Œ๋ฉด ๋‹ด๋ณด ๋น„์œจ์ด ๋จผ์ € ๋‚˜๋น ์ง„๋‹ค. ๊ทธ๋ž˜์„œ ํ•˜๋ฝ์žฅ์—์„œ๋Š” ์ฝ”์ธ์„ ๋” ํŒŒ๋Š” ์ชฝ์œผ๋กœ ๋ชฐ๋ฆฌ๊ธฐ ์‰ฝ๋‹ค. ์ด๋Ÿฐ ๋งค๋„๋Š” ๊ฐœ๋ณ„ ํšŒ์‚ฌ์˜ ๋ฌธ์ œ๋กœ ๋๋‚˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ๋‹ค. ์—์ด๋ธŒ ๊ฐ™์€ ํ”„๋กœํ† ์ฝœ์˜ ์ž๋™ ์ฒญ์‚ฐ์€ ์ •ํ•ด์ง„ ๊ฐ€๊ฒฉ์„ ์—์„œ ๊ธฐ๊ณ„์ ์œผ๋กœ ์ผ์–ด๋‚˜๊ธฐ ๋•Œ๋ฌธ์— ๋น„์Šทํ•œ ๊ตฌ์กฐ์˜ ๋Œ€์ถœ์ด ์—ฌ๋Ÿฌ ๊ณณ์— ์Œ“์—ฌ ์žˆ์œผ๋ฉด ํŠน์ • ๊ฐ€๊ฒฉ๋Œ€์—์„œ ๋งค๋ฌผ์ด ํ•œ๊บผ๋ฒˆ์— ๋‚˜์˜ฌ ์ˆ˜ ์žˆ๋‹ค. ํŠธ๋ ˆ์ €๋ฆฌ ๊ธฐ์—…์„ ๋ณผ ๋•Œ๋Š” ๋ณด์œ  ์ฝ”์ธ ์ˆ˜๋งŒํผ ๊ทธ ์ฝ”์ธ์ด ์–ด๋””์— ์–ผ๋งˆ๋‚˜ ๋ฌถ์—ฌ ์žˆ๋Š”์ง€๋ฅผ ๊ฐ™์ด ๋ด์•ผ ํ•œ๋‹ค. ๋‘ ํšŒ์‚ฌ์˜ ๊ณต์‹œ๋Š” ๋‹ด๋ณด๋กœ ์žกํžŒ ๋น„์ค‘๊ณผ ์ฒญ์‚ฐ์„ , ๋งŒ๊ธฐ์™€ ํ˜„๊ธˆ ์—ฌ์œ ๊ฐ€ ํ•˜๋ฝ์žฅ์—์„œ ํšŒ์‚ฌ์˜ ์„ ํƒ์ง€๋ฅผ ์ •ํ•œ๋‹ค๋Š” ์ ์„ ๋ณด์—ฌ ์ค€๋‹ค.
2
2
126
๐Ÿ“š Rialo Research Series 61: Trust Recovery - Can Autonomous Agents Unlearn Manipulated Evidence? (Kor : ์‹ ๋ขฐ ํšŒ๋ณต - ์ž์œจ Agent๋Š” ์กฐ์ž‘๋œ ์ฆ๊ฑฐ์˜ ์˜ํ–ฅ์„ ์ œ๊ฑฐํ•  ์ˆ˜ ์žˆ์„๊นŒ?) 1/ Intro: When Trust Becomes Corrupted (EN) In the previous chapter, we explored how attackers could manipulate an agent's confidence in its data sources. But detecting manipulation doesn't automatically repair the damage. Imagine an accurate oracle losing its reputation after several attacker-controlled feeds falsely identify it as unreliable. Even after the attack ends, the system may continue ignoring the trustworthy source. Stopping manipulation and recovering trust are two different problems. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ๊ณต๊ฒฉ์ž๊ฐ€ Oracle์˜ Confidence Score๋ฅผ ์กฐ์ž‘ํ•ด Agent์˜ ํŒ๋‹จ์„ ์™œ๊ณกํ•˜๋Š” ์ƒํ™ฉ์„ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ๊ณต๊ฒฉ์„ ๋ฐœ๊ฒฌํ–ˆ๋‹ค๊ณ  ํ•ด์„œ ๋ฌธ์ œ๊ฐ€ ๋ฐ”๋กœ ํ•ด๊ฒฐ๋˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ์ •์ƒ์ ์ธ Oracle์ด ๋‚ฎ์€ ์‹ ๋ขฐ๋„๋ฅผ ๋ฐ›์€ ์ƒํƒœ๋ผ๋ฉด ๊ณต๊ฒฉ์ด ๋๋‚œ ๋’ค์—๋„ ๊ณ„์† ๋ฌด์‹œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์กฐ์ž‘์„ ์ค‘๋‹จํ•˜๋Š” ๊ฒƒ๊ณผ ์˜ค์—ผ๋œ ์‹ ๋ขฐ๋ฅผ ํšŒ๋ณตํ•˜๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. 2/ Why Resetting Scores Isn't Enough (EN) One solution is resetting all reputation scores. But that erases legitimate performance history alongside manipulated evaluations. An attacker-controlled source could regain the same standing as a historically reliable oracle. A better recovery process identifies which observations were compromised and reassesses their influence. (KR) ๋ชจ๋“  Reputation Score๋ฅผ ์ดˆ๊ธฐํ™”ํ•˜๋ฉด ๊ฐ„๋‹จํ•ด ๋ณด์ด์ง€๋งŒ ์ •์ƒ์ ์ธ ํ‰๊ฐ€ ๊ธฐ๋ก๊นŒ์ง€ ์‚ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์ž๊ฐ€ ํ†ต์ œํ•˜๋˜ Oracle๋„ ์˜ค๋žซ๋™์•ˆ ์•ˆ์ •์ ์œผ๋กœ ์šด์˜๋œ Oracle๊ณผ ๊ฐ™์€ ์ ์ˆ˜์—์„œ ๋‹ค์‹œ ์‹œ์ž‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ „์ฒด ์ดˆ๊ธฐํ™”๋ณด๋‹ค ์กฐ์ž‘๋œ ํ‰๊ฐ€ ๊ธฐ๋ก์„ ์‹๋ณ„ํ•˜๊ณ  ์˜ํ–ฅ์„ ์žฌ๊ณ„์‚ฐํ•˜๋Š” ๋ฐฉ์‹์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 3/ Evidence Provenance (EN) Recovery depends on knowing where each confidence adjustment originated. Suppose an oracle lost 20 reputation points after five disagreement events. If three events relied on manipulated feeds, the application should be able to identify them. Recording evidence sources, timestamps, and policy versions makes targeted reassessment possible. (KR) ์‹ ๋ขฐ๋„๋ฅผ ๋ณต๊ตฌํ•˜๋ ค๋ฉด ์ ์ˆ˜๊ฐ€ ๋ณ€๊ฒฝ๋œ ๊ทผ๊ฑฐ๋ฅผ ์ถ”์ ํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. Oracle์ด ๋‹ค์„ฏ ๋ฒˆ์˜ ๊ฐ€๊ฒฉ ๋ถˆ์ผ์น˜๋กœ 20์ ์„ ์žƒ์—ˆ๋Š”๋ฐ ๊ทธ์ค‘ ์„ธ ๋ฒˆ์ด ์กฐ์ž‘๋œ ๋ฐ์ดํ„ฐ ๋•Œ๋ฌธ์ด์—ˆ๋‹ค๋ฉด ์–ด๋–จ๊นŒ์š”? Evidence Source, Timestamp, Policy Version์„ ๊ธฐ๋กํ•ด ๋‘๋ฉด ๋ฌธ์ œ๊ฐ€ ๋œ ํ‰๊ฐ€๋งŒ ๋‹ค์‹œ ๊ฒ€ํ† ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๋ขฐ ํšŒ๋ณต์˜ ์ถœ๋ฐœ์ ์€ ํ‰๊ฐ€ ์ด๋ ฅ์„ ์ถ”์ ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. 4/ Rebuilding Confidence Gradually (EN) Immediately restoring a source to its previous reputation can also be risky. The source may have been compromised independently of the detected attack. A recovery policy could require fresh independent observations before increasing its score, with limits on how quickly confidence changes. This reduces the risk of another sudden trust reversal. (KR) ์กฐ์ž‘์ด ํ™•์ธ๋๋‹ค๊ณ  ์ด์ „ ์‹ ๋ขฐ๋„๋ฅผ ์ฆ‰์‹œ ๋ณต์›ํ•˜๋Š” ๊ฒƒ๋„ ์œ„ํ—˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•ด๋‹น Oracle์— ๋ณ„๋„์˜ ๋ฌธ์ œ๊ฐ€ ์กด์žฌํ•  ๊ฐ€๋Šฅ์„ฑ๊นŒ์ง€ ๋ฐฐ์ œํ•  ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋…๋ฆฝ์ ์ธ ์ตœ์‹  ๋ฐ์ดํ„ฐ๋ฅผ ์ผ์ • ๊ธฐ๊ฐ„ ํ™•์ธํ•˜๊ณ  ์ ์ˆ˜๋ฅผ ๋‹จ๊ณ„์ ์œผ๋กœ ํšŒ๋ณตํ•˜๋Š” ๋ฐฉ์‹์„ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹ ๋ขฐ ํšŒ๋ณต์—๋„ ๊ฒ€์ฆ ๊ณผ์ •๊ณผ ์†๋„ ์ œํ•œ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 5/ The Feedback Loop Problem (EN) A low-ranked oracle may receive little influence in future decisions, making it difficult to demonstrate renewed reliability. This creates a recovery trap: the system distrusts the source because of its score, then denies it opportunities to improve that score. Independent evaluation can help break this loop without immediately granting full operational influence. (KR) ๋‚ฎ์€ ์ ์ˆ˜๋ฅผ ๋ฐ›์€ Oracle์€ ์ดํ›„ ํŒ๋‹จ์—์„œ ์˜ํ–ฅ๋ ฅ์ด ์ค„์–ด๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์˜ํ–ฅ๋ ฅ์ด ๋‚ฎ๋‹ค๋Š” ์ด์œ ๋กœ ํ‰๊ฐ€ ๊ธฐํšŒ๊นŒ์ง€ ์ค„์–ด๋“ ๋‹ค๋ฉด ์‹ ๋ขฐ๋„๋ฅผ ํšŒ๋ณตํ•˜๊ธฐ ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค. ๊ณผ๊ฑฐ์˜ ์ž˜๋ชป๋œ ํ‰๊ฐ€๊ฐ€ ๋ฏธ๋ž˜์˜ ํšŒ๋ณต๊นŒ์ง€ ๋ฐฉํ•ดํ•˜๋Š” ๊ตฌ์กฐ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ๋ฐฉ์ง€ํ•˜๋ ค๋ฉด ์‹ค์ œ ๊ฑฐ๋ž˜์— ์ ์šฉํ•˜๋Š” Weight์™€ ๋ณ„๋„๋กœ ๋…๋ฆฝ์ ์ธ ํ‰๊ฐ€ ๊ณผ์ •์„ ์šด์˜ํ•˜๋Š” ๋ฐฉ๋ฒ•๋„ ์žˆ์Šต๋‹ˆ๋‹ค. 6/ Rialo and Reactive Recovery (EN) Rialo's Reactive Transactions model allows defined conditions to trigger transactions during block execution. An application could use this mechanism to initiate restrictions when confidence falls and controlled recovery actions when independent evidence improves. However, Rialo does not automatically identify poisoned observations or calculate corrected reputation scores. Those responsibilities remain with application logic and its governance rules. (KR) Rialo์˜ Reactive Transactions๋ฅผ ํ™œ์šฉํ•˜๋ฉด Confidence ํ•˜๋ฝ ์‹œ ์ œํ•œ์„ ์‹คํ–‰ํ•˜๊ณ , ๋…๋ฆฝ์ ์ธ ๊ฒ€์ฆ ๊ฒฐ๊ณผ๊ฐ€ ๊ฐœ์„ ๋˜๋ฉด ๋ณต๊ตฌ ์ž‘์—…์„ Triggerํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ๊ตฌ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Rialo๊ฐ€ ์กฐ์ž‘๋œ ํ‰๊ฐ€๋ฅผ ์ž๋™์œผ๋กœ ์‹๋ณ„ํ•˜๊ฑฐ๋‚˜ Reputation Score๋ฅผ ์žฌ๊ณ„์‚ฐํ•ด ์ฃผ๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๋ณต๊ตฌ ์กฐ๊ฑด๊ณผ ๊ถŒํ•œ, ์ ์ˆ˜ ์กฐ์ • ๋ฐฉ์‹์€ Application์—์„œ ์ •์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a reactive lending demonstration where manipulated oracle feeds damage an accurate source's reputation. After identifying the attack, the application should isolate compromised observations, recalculate confidence, and gradually restore the source's influence. The demonstration should also show how the system prevents an attacker from exploiting the recovery process itself. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ Demo๋Š” ์กฐ์ž‘๋œ Oracle ๋ฐ์ดํ„ฐ ๋•Œ๋ฌธ์— ์ •์ƒ์ ์ธ Source์˜ Reputation์ด ํ•˜๋ฝํ•˜๋Š” ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์„ ๋ฐœ๊ฒฌํ•œ ๋’ค ์˜ค์—ผ๋œ ํ‰๊ฐ€๋ฅผ ๋ถ„๋ฆฌํ•˜๊ณ  Confidence๋ฅผ ์žฌ๊ณ„์‚ฐํ•˜๋ฉฐ ์‹ค์ œ ๊ฑฐ๋ž˜์— ๋ฐ˜์˜๋˜๋Š” Weight๋ฅผ ๋‹จ๊ณ„์ ์œผ๋กœ ๋ณต๊ตฌํ•˜๋Š” ๊ณผ์ •์„ ํ™•์ธํ•˜๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ๋ณต๊ตฌ ์ ˆ์ฐจ ์ž์ฒด๊ฐ€ ์ƒˆ๋กœ์šด ๊ณต๊ฒฉ ๊ฒฝ๋กœ๊ฐ€ ๋˜์ง€ ์•Š๋Š”์ง€๋„ ์ค‘์š”ํ•˜๊ฒ ์ฃ . ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ž๋™ํ™”๋Š” ์ž˜๋ชป๋œ ํŒ๋‹จ์—์„œ ํšŒ๋ณตํ•˜๋Š” ๋Šฅ๋ ฅ๊นŒ์ง€ ๊ฐ–์ถฐ์•ผ ํ•ฉ๋‹ˆ๋‹ค. (EN) We've explored why recovering trust requires evidence tracing, selective reassessment, and controlled reputation updates. But if an agent has already made financial decisions using manipulated confidence scores, can correcting its reputation model truly repair the consequences? (KR) ์ด๋ฒˆ ํŽธ์—์„œ๋Š” Trust Recovery๋ฅผ ์ค‘์‹ฌ์œผ๋กœ ์กฐ์ž‘๋œ ํ‰๊ฐ€ ๊ธฐ๋ก์„ ์ถ”์ ํ•˜๊ณ  ์‹ ๋ขฐ๋„๋ฅผ ๋ณต๊ตฌํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์ด๋ฏธ ์ž˜๋ชป๋œ Confidence Score๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์ฐจ์ž…์ด๋‚˜ ์ฒญ์‚ฐ์ด ์‹คํ–‰๋๋‹ค๋ฉด ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ• ๊นŒ์š”? ์‹ ๋ขฐ๋„ ์ ์ˆ˜๋ฅผ ์ˆ˜์ •ํ•˜๋Š” ๊ฒƒ๊ณผ ๊ทธ๋™์•ˆ ๋ฐœ์ƒํ•œ ๊ธˆ์œต ๊ฑฐ๋ž˜์˜ ๊ฒฐ๊ณผ๋ฅผ ๋˜๋Œ๋ฆฌ๋Š” ๊ฒƒ์€ ์ „ํ˜€ ๋‹ค๋ฅธ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. Autonomous Finance์—์„œ๋Š” ์ž˜๋ชป๋œ ํŒ๋‹จ์„ ์ˆ˜์ •ํ•˜๋Š” ๊ฒƒ๋ฟ ์•„๋‹ˆ๋ผ ์ด๋ฏธ ๋ฐœ์ƒํ•œ ํ”ผํ•ด๋ฅผ ์–ด๋–ป๊ฒŒ ์ฒ˜๋ฆฌํ• ์ง€๋„ ๊ณ ๋ฏผํ•ด์•ผ ํ•˜์ง€ ์•Š์„๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
36
๋น„ํŠธ๊ฒŸ ์›”๋ ›์ด ์–ด์ œ ์ €๋… 3๋ถ„๊ธฐ ํ”ผ์‹ฑ ๊ฒฝ๋ณด ์ˆ˜์น˜๋ฅผ ๊ณต๊ฐœํ–ˆ๋Š”๋ฐ ํ”ผ์‹ฑ ์˜์‹ฌ์œผ๋กœ ๊ฑธ๋Ÿฌ ๋‚ธ ์„œ๋ช… ์š”์ฒญ์ด ์ฃผ๋‹น ํ‰๊ท  3๋งŒ 7์ฒœ ๊ฑด์œผ๋กœ ์ง์ „ ๋ถ„๊ธฐ๋ณด๋‹ค 7๋ฐฐ ๋„˜๊ฒŒ ๋Š˜์—ˆ๋‹ค๊ณ  ํ•œ๋‹ค. ์ด ์ง€๊ฐ‘์„ ์“ฐ์ง€ ์•Š๋”๋ผ๋„ ์ง€๊ฐ‘์—์„œ ์„œ๋ช… ๋ฒ„ํŠผ์„ ๋ˆ„๋ฅด๋Š” ์‚ฌ๋žŒ์ด๋ผ๋ฉด ํ•œ ๋ฒˆ ๋ณผ ๋งŒํ•œ ์ˆซ์ž์ด๋‹ค. ์ด ์ˆซ์ž๋ฅผ ๊ทธ๋Œ€๋กœ ๊ณต๊ฒฉ ์ฆ๊ฐ€๋กœ ์ฝ๋Š” ๊ฑด ์กฐ์‹ฌํ•ด์•ผ ํ•œ๋‹ค. ๋น„ํŠธ๊ฒŸ ์›”๋ › ์ตœ๊ณ ์šด์˜์ฑ…์ž„์ž ์•จ๋นˆ ์นธ์€ Cointelegraph์— ๊ณต๊ฒฉ ์ž์ฒด๊ฐ€ ๋Š˜์–ด๋‚œ ์˜ํ–ฅ๊ณผ ํƒ์ง€ ๋Šฅ๋ ฅ์ด ์ข‹์•„์ง„ ์˜ํ–ฅ์ด ์„ž์—ฌ ์žˆ๊ณ  ๋‘˜์„ ๋‚˜๋ˆ  ์…€ ์ˆ˜๋Š” ์—†๋‹ค๊ณ  ๋งํ–ˆ๋‹ค. ๊ฒฝ๋ณด ๊ฑด์ˆ˜๋Š” ๊ฑธ๋Ÿฌ ๋‚ธ ์„œ๋ช… ์š”์ฒญ์˜ ์ˆ˜๋ผ์„œ ์‹ค์ œ ๊ณต๊ฒฉ ๊ฑด์ˆ˜๋‚˜ ํ™•์ธ๋œ ํ”ผํ•ด์•ก๊ณผ๋„ ๋‹ค๋ฅด๋‹ค. ๊ทธ๋ณด๋‹ค ๋ˆˆ์—ฌ๊ฒจ๋ณผ ๊ฑด ์ฐจ๋‹จ ๋„๋ฉ”์ธ ์ชฝ ์ˆ˜์น˜์ด๋‹ค. ๋น„ํŠธ๊ฒŸ ์›”๋ ›์€ ์ง€๋‚œ๋‹ฌ ์ƒˆ๋กœ ๋ง‰์€ ์œ„ํ—˜ ๋„๋ฉ”์ธ ๊ฐ€์šด๋ฐ 90%๋ฅผ AI ๊ธฐ๋ฐ˜ ํƒ์ง€๊ฐ€ ์ฐพ์•„๋ƒˆ๊ณ  ๊ทธ์ค‘ 70% ๊ฐ€๊นŒ์ด๋Š” ์™ธ๋ถ€ ์œ„ํ˜‘ ๋ชฉ๋ก์— ์•„์ง ์˜ฌ๋ผ์™€ ์žˆ์ง€ ์•Š๋˜ ๊ณณ์ด์—ˆ๋‹ค๊ณ  ๋ฐํ˜”๋‹ค. ๊ธฐ์กด ๋ธ”๋ž™๋ฆฌ์ŠคํŠธ๋งŒ ๋ฏฟ๊ณ  ์žˆ์œผ๋ฉด ์ƒˆ๋กœ ๋งŒ๋“ค์–ด์ง„ ํ”ผ์‹ฑ ์‚ฌ์ดํŠธ๋Š” ํ•œ๋™์•ˆ ๊ทธ๋Œ€๋กœ ํ†ต๊ณผํ•œ๋‹ค๋Š” ๋œป์œผ๋กœ ์ฝํžŒ๋‹ค. ์ง€๊ฐ‘ ํšŒ์‚ฌ์˜ ํƒ์ง€๊ฐ€ ์ข‹์•„์ง€๋Š” ๊ฑด ๋ฐ˜๊ฐ€์šด ์ผ์ด๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ๊ฒฐ๊ตญ ๋งˆ์ง€๋ง‰ ํ™•์ธ์€ ์„œ๋ช…ํ•˜๋Š” ์‚ฌ๋žŒ์ด ํ•ด์•ผ ํ•˜๋Š” ๋งŒํผ ์Šน์ธ ์š”์ฒญ์ด ๋œจ๋ฉด ์–ด๋–ค ์ปจํŠธ๋ž™ํŠธ์— ์–ด๋–ค ๊ถŒํ•œ์„ ์ฃผ๋Š”์ง€๋ถ€ํ„ฐ ๋ด์•ผ ํ•œ๋‹ค. ์ง€๊ฐ‘์ด ๊ฑฐ๋ž˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์„ ์ง€์›ํ•œ๋‹ค๋ฉด ๊ฒฐ๊ณผ๋ฅผ ํ•œ ๋ฒˆ ํ™•์ธํ•œ ๋’ค ๋ˆ„๋ฅด๋Š” ์Šต๊ด€๋„ ํ•„์š”ํ•˜๋‹ค. ์˜ค๋ž˜ ์“ฐ์ง€ ์•Š์€ ์Šน์ธ์€ ์ฃผ๊ธฐ์ ์œผ๋กœ ์ •๋ฆฌํ•ด ๋‘๋Š” ํŽธ์ด ์ข‹๋‹ค.
2
99
๐Ÿ“š Rialo Research Series 60: Trust Manipulation - Can Attackers Poison an Agent's Confidence? (Kor : ์‹ ๋ขฐ๋„ ์กฐ์ž‘ - ๊ณต๊ฒฉ์ž๊ฐ€ Agent์˜ ์‹ ๋ขฐ ํŒ๋‹จ์„ ์˜ค์—ผ์‹œํ‚ฌ ์ˆ˜ ์žˆ์„๊นŒ?) 1/ Intro: Attacking Trust Instead of Data (EN) In the previous chapter, we explored how agents evaluate conflicting evidence. But what if attackers target the evaluation process itself? Imagine a lending protocol assigning confidence scores to multiple oracles. An attacker cannot directly change their prices, so instead attempts to make a reliable oracle appear untrustworthy. Manipulating what an agent trusts can be as dangerous as manipulating the data it receives. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ์—ฌ๋Ÿฌ ๋ฐ์ดํ„ฐ ์†Œ์Šค๊ฐ€ ์ถฉ๋Œํ•  ๋•Œ ์–ด๋–ค ์ฆ๊ฑฐ๋ฅผ ์‹ ๋ขฐํ• ์ง€ ํŒ๋‹จํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ๊ณต๊ฒฉ์ž๊ฐ€ ๋ฐ์ดํ„ฐ๊ฐ€ ์•„๋‹ˆ๋ผ ์‹ ๋ขฐ๋„ ํ‰๊ฐ€ ๊ณผ์ • ์ž์ฒด๋ฅผ ๋…ธ๋ฆฐ๋‹ค๋ฉด ์–ด๋–จ๊นŒ์š”? Lending Protocol์ด ์—ฌ๋Ÿฌ Oracle์— Confidence Score๋ฅผ ๋ถ€์—ฌํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์ž๋Š” ๊ฐ€๊ฒฉ์„ ์ง์ ‘ ์กฐ์ž‘ํ•˜๋Š” ๋Œ€์‹  ์ •์ƒ์ ์ธ Oracle์˜ ์‹ ๋ขฐ๋„๋ฅผ ๋–จ์–ด๋œจ๋ฆฌ๋ ค๊ณ  ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ”๊พธ์ง€ ์•Š์•„๋„ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ์‹ ๋ขฐ๋ฅผ ์กฐ์ž‘ํ•˜๋ฉด ๊ฒฐ๊ณผ์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 2/ Reputation Poisoning (EN) Suppose an oracle's confidence score decreases whenever its price differs significantly from the majority. An attacker temporarily manipulates several low-quality feeds, making an accurate oracle appear to be the outlier. If the system automatically penalizes disagreement, the trustworthy source could lose influence. This is a form of reputation poisoning through manipulated comparisons. (KR) ๋‹ค์ˆ˜์˜ Oracle๊ณผ ๊ฐ€๊ฒฉ ์ฐจ์ด๊ฐ€ ํฌ๋ฉด Confidence Score๊ฐ€ ๊ฐ์†Œํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ์ƒ๊ฐํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์ž๊ฐ€ ํ’ˆ์งˆ์ด ๋‚ฎ์€ ์—ฌ๋Ÿฌ ๋ฐ์ดํ„ฐ ์†Œ์Šค๋ฅผ ์ผ์‹œ์ ์œผ๋กœ ์กฐ์ž‘ํ•˜๋ฉด ์ •ํ™•ํ•œ Oracle์ด ์˜คํžˆ๋ ค ์ด์ƒ๊ฐ’์ฒ˜๋Ÿผ ๋ณด์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์‹œ์Šคํ…œ์ด ๋ถˆ์ผ์น˜๋งŒ์„ ๊ธฐ์ค€์œผ๋กœ ์ ์ˆ˜๋ฅผ ๋‚ฎ์ถ˜๋‹ค๋ฉด ์ •์ƒ์ ์ธ Oracle์˜ ์˜ํ–ฅ๋ ฅ์ด ๊ฐ์†Œํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์ˆ˜์™€ ๋‹ค๋ฅด๋‹ค๋Š” ์ด์œ ๋งŒ์œผ๋กœ ์‹ ๋ขฐ๋„๋ฅผ ๋‚ฎ์ถ”๋Š” ๋ฐฉ์‹์€ ๊ณต๊ฒฉ์— ์•…์šฉ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 3/ Slow Trust Manipulation (EN) Trust attacks do not always require sudden price changes. An attacker might gradually influence source rankings through repeated small deviations or carefully timed anomalies. Each event appears minor, but the accumulated score changes may eventually favor attacker-controlled sources. Confidence updates therefore need limits, observation windows, and safeguards against coordinated manipulation. (KR) ์‹ ๋ขฐ๋„ ์กฐ์ž‘์ด ๋ฐ˜๋“œ์‹œ ๊ธ‰๊ฒฉํ•œ ๊ฐ€๊ฒฉ ๋ณ€๋™์œผ๋กœ ๋ฐœ์ƒํ•˜๋Š” ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. ๊ณต๊ฒฉ์ž๋Š” ์ž‘์€ ๊ฐ€๊ฒฉ ์ฐจ์ด๋‚˜ ํŠน์ • ์‹œ์ ์˜ ์ด์ƒ ์‹ ํ˜ธ๋ฅผ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋งŒ๋“ค์–ด Source Ranking์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐœ๋ณ„ ์‚ฌ๊ฑด์€ ์‚ฌ์†Œํ•ด ๋ณด์—ฌ๋„ ๋ˆ„์ ๋œ ์ ์ˆ˜ ๋ณ€ํ™”๊ฐ€ ๊ณต๊ฒฉ์ž์—๊ฒŒ ์œ ๋ฆฌํ•˜๊ฒŒ ์ž‘์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Confidence Update์—๋Š” ๋ณ€๊ฒฝ ํญ ์ œํ•œ๊ณผ ๊ด€์ฐฐ ๊ธฐ๊ฐ„, ์กฐ์ง์ ์ธ ์กฐ์ž‘์„ ๋ฐฉ์ง€ํ•˜๋Š” ๊ธฐ์ค€์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. 4/ Feedback Loops (EN) A dangerous feedback loop emerges when source rankings influence future evaluations. A low-ranked oracle receives less weight, so its accurate warnings may be ignored. Meanwhile, highly ranked sources reinforce one another even when sharing faulty upstream data. Over time, the system can become increasingly confident in an incorrect consensus. Confidence is not proof of correctness. (KR) Source Ranking์ด ๋‹ค์Œ ํ‰๊ฐ€์—๋„ ์˜ํ–ฅ์„ ์ค€๋‹ค๋ฉด Feedback Loop๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‚ฎ์€ ์ ์ˆ˜๋ฅผ ๋ฐ›์€ Oracle์€ ์˜ํ–ฅ๋ ฅ์ด ์ค„์–ด๋“ค์–ด ์ •ํ™•ํ•œ ๊ฒฝ๊ณ ๋ฅผ ์ œ๊ณตํ•ด๋„ ๋ฌด์‹œ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋Œ€๋กœ ๋†’์€ ์ ์ˆ˜๋ฅผ ๊ฐ€์ง„ Oracle๋“ค์€ ๋™์ผํ•œ ์˜ค๋ฅ˜๋ฅผ ๊ณต์œ ํ•˜๋ฉด์„œ ์„œ๋กœ์˜ ํŒ๋‹จ์„ ๊ฐ•ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ ์‹œ์Šคํ…œ์ด ์ž˜๋ชป๋œ ๊ฒฐ๋ก ์— ์ ์  ๋” ๋†’์€ ํ™•์‹ ์„ ๊ฐ–๊ฒŒ ๋˜๋Š” ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. Confidence Score๊ฐ€ ๋†’๋‹ค๊ณ  ์‹ค์ œ ํŒ๋‹จ๊นŒ์ง€ ์ •ํ™•ํ•œ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. 5/ Protecting Trust Updates (EN) One defense is separating immediate risk detection from long-term reputation updates. A single anomaly might trigger temporary restrictions without permanently damaging an oracle's score. Reputation changes could require independent evidence, capped adjustments, and periodic review. The goal is to prevent short-lived incidents from rewriting the system's long-term trust assumptions. (KR) ํ•œ ๊ฐ€์ง€ ๋Œ€์‘ ๋ฐฉ๋ฒ•์€ ์ฆ‰๊ฐ์ ์ธ ์œ„ํ—˜ ๊ฐ์ง€์™€ ์žฅ๊ธฐ์ ์ธ Reputation Update๋ฅผ ๋ถ„๋ฆฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ผ์‹œ์ ์ธ ์ด์ƒ ์‹ ํ˜ธ๋Š” ๊ฑฐ๋ž˜ ์ œํ•œ์„ Triggerํ•  ์ˆ˜ ์žˆ์ง€๋งŒ Oracle์˜ ์žฅ๊ธฐ ์‹ ๋ขฐ๋„๊นŒ์ง€ ์ฆ‰์‹œ ๋‚ฎ์ถœ ํ•„์š”๋Š” ์—†์Šต๋‹ˆ๋‹ค. ์‹ ๋ขฐ๋„ ๋ณ€๊ฒฝ์—๋Š” ๋…๋ฆฝ์ ์ธ ์ฆ๊ฑฐ์™€ ์กฐ์ • ํญ ์ œํ•œ, ์ •๊ธฐ์ ์ธ ๊ฒ€ํ†  ์ ˆ์ฐจ๋ฅผ ์ ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹จ๊ธฐ์ ์ธ ์ด์ƒ ํ˜„์ƒ์ด ์žฅ๊ธฐ์ ์ธ ์‹ ๋ขฐ ๊ธฐ์ค€์„ ์‰ฝ๊ฒŒ ๋ฐ”๊พธ์ง€ ๋ชปํ•˜๋„๋ก ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 6/ Rialo and Reactive Trust Controls (EN) Rialo's Reactive Transactions model allows defined conditions to trigger associated transactions during block execution. Applications could use this capability to restrict borrowing when oracle confidence falls below a threshold. However, Rialo does not automatically calculate trustworthy reputation scores or protect them against manipulation. Score updates, evidence validation, and recovery policies remain application-level responsibilities. (KR) Rialo์˜ Reactive Transactions๋Š” ์ •์˜๋œ ์กฐ๊ฑด์— ๋”ฐ๋ผ ๊ด€๋ จ Transaction์„ Triggerํ•˜๋Š” ๊ตฌ์กฐ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ํ™œ์šฉํ•˜๋ฉด Oracle Confidence๊ฐ€ ์ผ์ • ๊ธฐ์ค€ ์•„๋ž˜๋กœ ๋–จ์–ด์กŒ์„ ๋•Œ ์‹ ๊ทœ ์ฐจ์ž…์„ ์ œํ•œํ•˜๋Š” Application์„ ์„ค๊ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Rialo๊ฐ€ ์‹ ๋ขฐ๋„ ์ ์ˆ˜๋ฅผ ์ž๋™์œผ๋กœ ๊ณ„์‚ฐํ•˜๊ฑฐ๋‚˜ ์กฐ์ž‘์„ ๋ฐฉ์ง€ํ•ด ์ค€๋‹ค๋Š” ์˜๋ฏธ๋Š” ์•„๋‹™๋‹ˆ๋‹ค. Score Update, Evidence Validation, Recovery Policy๋Š” Application์—์„œ ์ง์ ‘ ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a reactive lending demonstration where an attacker manipulates several low-quality oracle feeds to discredit an accurate source. The application should compare immediate safety responses with longer-term reputation adjustments. It should also demonstrate whether the accurate oracle can recover its score once the manipulation is detected. This would reveal how an automated system protects not just its decisions, but the trust assumptions behind them. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ Demo๋Š” ๊ณต๊ฒฉ์ž๊ฐ€ ์—ฌ๋Ÿฌ ์ €ํ’ˆ์งˆ Oracle์„ ์กฐ์ž‘ํ•ด ์ •ํ™•ํ•œ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ์‹ ๋ขฐ๋„๋ฅผ ๋–จ์–ด๋œจ๋ฆฌ๋Š” ์ƒํ™ฉ์ž…๋‹ˆ๋‹ค. Application์ด ์ฆ‰๊ฐ์ ์ธ Safety Response์™€ ์žฅ๊ธฐ์ ์ธ Reputation Adjustment๋ฅผ ์–ด๋–ป๊ฒŒ ๊ตฌ๋ถ„ํ•˜๋Š”์ง€ ํ™•์ธํ•˜๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ์กฐ์ž‘์ด ๋ฐœ๊ฒฌ๋œ ๋’ค ์ •์ƒ์ ์ธ Oracle์ด ์‹ ๋ขฐ๋„๋ฅผ ํšŒ๋ณตํ•  ์ˆ˜ ์žˆ๋Š”์ง€๋„ ์ค‘์š”ํ•˜๊ฒ ์ฃ . ์ž๋™ํ™”๋œ ์‹œ์Šคํ…œ์€ ํŒ๋‹จ ๊ฒฐ๊ณผ๋ฟ ์•„๋‹ˆ๋ผ ํŒ๋‹จ์˜ ๊ธฐ์ค€์ด ๋˜๋Š” ์‹ ๋ขฐ ๊ตฌ์กฐ๊นŒ์ง€ ๋ณดํ˜ธํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. (EN) We've explored how attackers might manipulate confidence scores without directly altering the underlying data. If an autonomous system learns which sources to trust from its own past decisions, how can it prevent earlier mistakes from shaping future judgments? (KR) ์ด๋ฒˆ ํŽธ์—์„œ๋Š” Trust Manipulation์„ ์ค‘์‹ฌ์œผ๋กœ ๊ณต๊ฒฉ์ž๊ฐ€ ๋ฐ์ดํ„ฐ ์ž์ฒด๋ฅผ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ ๋„ Confidence Score์™€ Source Ranking์„ ์กฐ์ž‘ํ•  ๊ฐ€๋Šฅ์„ฑ์„ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ Agent๊ฐ€ ๊ณผ๊ฑฐ ํŒ๋‹จ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์‹ ๋ขฐ ๊ธฐ์ค€์„ ๊ณ„์† ํ•™์Šตํ•œ๋‹ค๋ฉด ์–ด๋–จ๊นŒ์š”? ์ดˆ๊ธฐ์— ์ž˜๋ชป๋œ ํ‰๊ฐ€๊ฐ€ ๋ˆ„์ ๋˜๋ฉด์„œ ์ดํ›„์˜ ํŒ๋‹จ๊นŒ์ง€ ์™œ๊ณก๋  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. Autonomous Finance์—์„œ ์Šค์Šค๋กœ ํ•™์Šตํ•˜๋Š” ์‹ ๋ขฐ ์‹œ์Šคํ…œ์€ ์–ด๋–ป๊ฒŒ ๊ณผ๊ฑฐ์˜ ์ž˜๋ชป๋œ ํŒ๋‹จ์„ ์ˆ˜์ •ํ•˜๊ณ  ์ •์ƒ์ ์ธ ๊ธฐ์ค€์œผ๋กœ ๋Œ์•„๊ฐˆ ์ˆ˜ ์žˆ์„๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
๐Ÿ“š Rialo Research Series 59: Evidence Arbitration - When Data Sources Disagree, What Counts as Truth? (Kor : ์ฆ๊ฑฐ ์ค‘์žฌ - ๋ฐ์ดํ„ฐ ์†Œ์Šค๊ฐ€ ์ถฉ๋Œํ•  ๋•Œ ๋ฌด์—‡์„ ์ง„์‹ค๋กœ ํŒ๋‹จํ•ด์•ผ ํ• ๊นŒ?) 1/ Intro: When Evidence Contradicts Evidence (EN) In the previous chapter, we explored disagreements between independent verifiers. But what happens when the evidence itself conflicts? Imagine three price oracles reporting ETH at $2,000, $1,998, and $1,920. Two agree closely, while the third differs significantly. A lending agent must decide whether to approve a withdrawal. The difficult question isn't which answer is most popular, but which evidence is reliable enough to act upon. (KR) ์ง€๋‚œ ํŽธ์—์„œ๋Š” ๋…๋ฆฝ์ ์ธ Verifier๋“ค์ด ์„œ๋กœ ๋‹ค๋ฅธ ๊ฒฐ๋ก ์„ ๋‚ด๋ฆด ๋•Œ ๋ฐœ์ƒํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ๊ฒ€์ฆ์— ์‚ฌ์šฉํ•˜๋Š” ์ฆ๊ฑฐ ์ž์ฒด๊ฐ€ ์ถฉ๋Œํ•œ๋‹ค๋ฉด ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ• ๊นŒ์š”? ์„ธ ๊ฐœ์˜ Oracle์ด ETH ๊ฐ€๊ฒฉ์„ ๊ฐ๊ฐ 2,000๋‹ฌ๋Ÿฌ, 1,998๋‹ฌ๋Ÿฌ, 1,920๋‹ฌ๋Ÿฌ๋กœ ๋ณด๊ณ ํ–ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ๋‘ ๊ฐ€๊ฒฉ์€ ๋น„์Šทํ•˜์ง€๋งŒ ๋‚˜๋จธ์ง€ ํ•˜๋‚˜๋Š” ์ฐจ์ด๊ฐ€ ํฝ๋‹ˆ๋‹ค. Lending Agent๋Š” ์–ด๋–ค ๊ฐ€๊ฒฉ์„ ๊ธฐ์ค€์œผ๋กœ Withdrawal์„ ์Šน์ธํ•ด์•ผ ํ• ๊นŒ์š”? ์ค‘์š”ํ•œ ๊ฑด ๋‹ค์ˆ˜๊ฐ€ ์„ ํƒํ•œ ๊ฐ’์ด ์•„๋‹ˆ๋ผ ์‹ค์ œ๋กœ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ฆ๊ฑฐ๋ฅผ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. 2/ Majority Agreement Can Be Misleading (EN) A simple approach is to use the median price. In this example, that would be $1,998. But two oracles might depend on the same upstream exchange or data provider. If their shared source fails, apparent agreement can hide a common error. Source independence matters as much as the number of reports. (KR) ๊ฐ€์žฅ ๊ฐ„๋‹จํ•œ ๋ฐฉ๋ฒ•์€ Median Price๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์œ„ ์‚ฌ๋ก€์—์„œ๋Š” 1,998๋‹ฌ๋Ÿฌ๊ฐ€ ๊ธฐ์ค€์ด ๋ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋‘ Oracle์ด ๋™์ผํ•œ ๊ฑฐ๋ž˜์†Œ๋‚˜ ๋ฐ์ดํ„ฐ ๊ณต๊ธ‰์ž์— ์˜์กดํ•œ๋‹ค๋ฉด ์–ด๋–จ๊นŒ์š”? ๊ณตํ†ต ๋ฐ์ดํ„ฐ ์†Œ์Šค์— ์žฅ์• ๊ฐ€ ๋ฐœ์ƒํ•˜๋ฉด ๋‘ Oracle์ด ๊ฐ™์€ ์ž˜๋ชป๋œ ๊ฐ€๊ฒฉ์„ ์ œ๊ณตํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. Oracle์˜ ์ˆซ์ž๋ฟ ์•„๋‹ˆ๋ผ ๋ฐ์ดํ„ฐ ์ถœ์ฒ˜์˜ ๋…๋ฆฝ์„ฑ๋„ ํ™•์ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 3/ Freshness vs Accuracy (EN) The newest data is not necessarily the most accurate. Oracle A may update every second but rely on a thinly traded market. Oracle B may update less frequently while aggregating deeper liquidity. During volatility, their prices can diverge. An application should define acceptable data age, market quality, and deviation limits rather than blindly selecting the latest report. (KR) ๊ฐ€์žฅ ์ตœ์‹  ๋ฐ์ดํ„ฐ๊ฐ€ ๋ฐ˜๋“œ์‹œ ๊ฐ€์žฅ ์ •ํ™•ํ•œ ๊ฒƒ์€ ์•„๋‹™๋‹ˆ๋‹ค. Oracle A๋Š” 1์ดˆ๋งˆ๋‹ค ๊ฐ€๊ฒฉ์„ ๊ฐฑ์‹ ํ•˜์ง€๋งŒ ๊ฑฐ๋ž˜๋Ÿ‰์ด ์ ์€ ์‹œ์žฅ์„ ์ฐธ์กฐํ•˜๊ณ , Oracle B๋Š” ๊ฐฑ์‹ ์ด ๋А๋ฆฌ์ง€๋งŒ ์œ ๋™์„ฑ์ด ํ’๋ถ€ํ•œ ์—ฌ๋Ÿฌ ์‹œ์žฅ์„ ์ฐธ๊ณ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ณ€๋™์„ฑ์ด ์ปค์ง€๋ฉด ๋‘ ๊ฐ€๊ฒฉ์˜ ์ฐจ์ด๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ Timestamp๋ฟ ์•„๋‹ˆ๋ผ Liquidity, Data Quality, Deviation Threshold ๋“ฑ์„ ํ•จ๊ป˜ ๊ณ ๋ คํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 4/ Evidence Weighting (EN) One possible design assigns different weights to evidence sources based on predefined criteria. For example, an oracle with strong liquidity coverage and reliable historical performance might receive greater weight. However, weighting introduces governance risks. Who determines the scores, and how are they updated? A compromised weighting policy could influence decisions without manipulating the underlying prices. (KR) ์ฆ๊ฑฐ๋งˆ๋‹ค ์„œ๋กœ ๋‹ค๋ฅธ ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•˜๋Š” ๋ฐฉ์‹๋„ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์œ ๋™์„ฑ์ด ํ’๋ถ€ํ•œ ์‹œ์žฅ์„ ์ฐธ์กฐํ•˜๊ณ  ๊ณผ๊ฑฐ ๋ฐ์ดํ„ฐ ํ’ˆ์งˆ์ด ์šฐ์ˆ˜ํ•œ Oracle์— ๋” ๋†’์€ Weight๋ฅผ ๋ถ€์—ฌํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ƒˆ๋กœ์šด ๋ฌธ์ œ๊ฐ€ ์ƒ๊น๋‹ˆ๋‹ค. ๊ฐ€์ค‘์น˜๋Š” ๋ˆ„๊ฐ€ ๊ฒฐ์ •ํ•˜๊ณ  ์–ด๋–ค ๊ธฐ์ค€์œผ๋กœ ๋ณ€๊ฒฝํ• ๊นŒ์š”? Weighting Policy ์ž์ฒด๊ฐ€ ์กฐ์ž‘๋œ๋‹ค๋ฉด ๊ฐ€๊ฒฉ ๋ฐ์ดํ„ฐ๋ฅผ ์ง์ ‘ ๋ณ€๊ฒฝํ•˜์ง€ ์•Š๊ณ ๋„ ํŒ๋‹จ ๊ฒฐ๊ณผ์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 5/ When No Source Is Trustworthy (EN) Sometimes no available source meets the required confidence threshold. Forcing a price selection could expose the protocol to unnecessary risk. A lending application might temporarily block new borrowing while preserving operations that do not depend on disputed prices. The fallback should be defined before an incident, not improvised during one. (KR) ์–ด๋–ค ๋ฐ์ดํ„ฐ๋„ ์ถฉ๋ถ„ํžˆ ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๋Š” ์ƒํ™ฉ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋•Œ ๋ฌด๋ฆฌํ•˜๊ฒŒ ํ•˜๋‚˜์˜ ๊ฐ€๊ฒฉ์„ ์„ ํƒํ•˜๋ฉด Protocol์ด ๋ถˆํ•„์š”ํ•œ ์œ„ํ—˜์„ ๊ฐ์ˆ˜ํ•˜๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ์‹ ๊ทœ ์ฐจ์ž…์€ ์ œํ•œํ•˜๋˜ ๋…ผ์Ÿ ์ค‘์ธ ๊ฐ€๊ฒฉ์— ์˜์กดํ•˜์ง€ ์•Š๋Š” ์ž‘์—…์€ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฐ Fallback Policy๋Š” ์žฅ์• ๊ฐ€ ๋ฐœ์ƒํ•˜๊ธฐ ์ „์— ์ •์˜๋ผ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ํŒ๋‹จํ•  ์ˆ˜ ์—†๋Š” ์ƒํ™ฉ์—์„œ๋Š” ํŒ๋‹จ์„ ๋ณด๋ฅ˜ํ•˜๋Š” ๊ฒƒ๋„ ์ค‘์š”ํ•œ ์•ˆ์ „์žฅ์น˜์ž…๋‹ˆ๋‹ค. 6/ Rialo and Evidence-Based Automation (EN) Rialo's Reactive Transactions model evaluates defined conditions during block execution and can trigger associated transactions. Applications could use this structure to react when oracle disagreement exceeds a predefined threshold. For example, excessive price divergence could trigger a borrowing restriction. However, Rialo does not automatically determine which oracle is truthful or provide a universal evidence-arbitration mechanism. Source selection, weighting, and dispute rules remain application responsibilities. (KR) Rialo์˜ Reactive Transactions๋Š” Block Execution ๊ณผ์ •์—์„œ ์กฐ๊ฑด์„ ํ‰๊ฐ€ํ•˜๊ณ  ๊ด€๋ จ Transaction์„ Triggerํ•˜๋Š” ๊ตฌ์กฐ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ํ™œ์šฉํ•˜๋ฉด Oracle ๊ฐ„ ๊ฐ€๊ฒฉ ์ฐจ์ด๊ฐ€ ํŠน์ • ๊ธฐ์ค€์„ ์ดˆ๊ณผํ–ˆ์„ ๋•Œ ๋ณดํ˜ธ ์ž‘์—…์„ ์‹คํ–‰ํ•˜๋„๋ก ์„ค๊ณ„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด Price Divergence๊ฐ€ ํ—ˆ์šฉ ๋ฒ”์œ„๋ฅผ ๋ฒ—์–ด๋‚˜๋ฉด ์‹ ๊ทœ ์ฐจ์ž…์„ ์ œํ•œํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ Rialo๊ฐ€ ์–ด๋–ค Oracle์ด ์ •ํ™•ํ•œ์ง€ ์ž๋™์œผ๋กœ ํŒ๋‹จํ•˜๊ฑฐ๋‚˜ ๋ฒ”์šฉ Evidence Arbitration์„ ์ œ๊ณตํ•œ๋‹ค๋Š” ์˜๋ฏธ๋Š” ์•„๋‹™๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ ํƒ๊ณผ ๊ฐ€์ค‘์น˜, ์ถฉ๋Œ ํ•ด๊ฒฐ ๊ทœ์น™์€ Application์—์„œ ๊ตฌํ˜„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. 7/ What I'd Like to See From Rialo (EN) I'd like to see a reactive lending demonstration using three oracle sources with different update speeds and liquidity profiles. The application should test normal agreement, stale data, shared-source failure, and extreme price divergence. Each outcome should reveal which evidence was accepted, rejected, or considered insufficient. That would help distinguish automated execution from trustworthy automated judgment. (KR) ๊ฐœ์ธ์ ์œผ๋กœ Rialo์—์„œ ๋ณด๊ณ  ์‹ถ์€ Demo๋Š” ๊ฐฑ์‹  ์†๋„์™€ ์œ ๋™์„ฑ ์กฐ๊ฑด์ด ์„œ๋กœ ๋‹ค๋ฅธ Oracle ์„ธ ๊ฐœ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” Reactive Lending Application์ž…๋‹ˆ๋‹ค. ์ •์ƒ์ ์ธ ๊ฐ€๊ฒฉ ์ผ์น˜, Stale Data, ๊ณตํ†ต ๋ฐ์ดํ„ฐ ์†Œ์Šค ์žฅ์• , ๊ทน๋‹จ์ ์ธ ๊ฐ€๊ฒฉ ์ฐจ์ด๋ฅผ ๊ฐ๊ฐ ํ…Œ์ŠคํŠธํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์–ด๋–ค ๋ฐ์ดํ„ฐ๊ฐ€ ์ฑ„ํƒ๋˜๊ฑฐ๋‚˜ ์ œ์™ธ๋๋Š”์ง€, ์–ด๋–ค ์ƒํ™ฉ์—์„œ ํŒ๋‹จ์ด ๋ณด๋ฅ˜๋๋Š”์ง€ ํ™•์ธํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ž๋™์œผ๋กœ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ๊ณผ ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ํŒ๋‹จ์„ ๋‚ด๋ฆฐ๋‹ค๋Š” ๊ฒƒ์€ ๋ณ„๊ฐœ์˜ ๋ฌธ์ œ์ž…๋‹ˆ๋‹ค. (EN) We've explored why evidence arbitration requires more than selecting the majority price or the newest data. But if an autonomous system repeatedly changes which sources it trusts, could an attacker manipulate those trust adjustments to influence future decisions? (KR) ์ด๋ฒˆ ํŽธ์—์„œ๋Š” Evidence Arbitration์„ ์ค‘์‹ฌ์œผ๋กœ ๋ฐ์ดํ„ฐ ์†Œ์Šค๊ฐ€ ์ถฉ๋Œํ•  ๋•Œ ๊ณ ๋ คํ•ด์•ผ ํ•  ๋…๋ฆฝ์„ฑ, ์ •ํ™•์„ฑ, ๊ฐ€์ค‘์น˜, Fallback Policy๋ฅผ ์‚ดํŽด๋ดค์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์—ฌ๊ธฐ์„œ ๋˜ ๋‹ค๋ฅธ ๋ฌธ์ œ๊ฐ€ ์ƒ๊น๋‹ˆ๋‹ค. ์‹œ์Šคํ…œ์ด ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ์‹ ๋ขฐ๋„๋ฅผ ์ง€์†์ ์œผ๋กœ ์กฐ์ •ํ•œ๋‹ค๋ฉด ๊ณต๊ฒฉ์ž๊ฐ€ ๊ทธ ๊ณผ์ •์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ์ง€ ์•Š์„๊นŒ์š”? ํŠน์ • Oracle์ด ๋ฐ˜๋ณต์ ์œผ๋กœ ์ž˜๋ชป๋œ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์ฒ˜๋Ÿผ ๋ณด์ด๊ฒŒ ๋งŒ๋“ค๊ฑฐ๋‚˜, ์ •์ƒ์ ์ธ ๋ฐ์ดํ„ฐ ์†Œ์Šค์˜ ์‹ ๋ขฐ๋„๋ฅผ ์˜๋„์ ์œผ๋กœ ๋‚ฎ์ถ”๋Š” ์ƒํ™ฉ๋„ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฒฐ๊ตญ Autonomous Finance์—์„œ๋Š” ๋ฐ์ดํ„ฐ ์ž์ฒด๋ฟ ์•„๋‹ˆ๋ผ ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ์‹ ๋ขฐํ• ์ง€ ๊ฒฐ์ •ํ•˜๋Š” ๊ธฐ์ค€๊นŒ์ง€ ๋ณดํ˜ธํ•ด์•ผ ํ•˜์ง€ ์•Š์„๊นŒ์š”? @RialoHQ @RialoKorea @Ade_Adepoju @silverwave1000 #Rialo
2
3
131
๋ฉ”๊ฐ€์ด๋”(MegaETH)๋ฅผ ๋งŒ๋“  ์žฌ๋‹จ์ด ์˜ค๋Š˜ ์ƒˆ๋ฒฝ ์ด๋ฆ„์„ MEGA๋กœ ๋ฐ”๊พธ๊ณ  ์žฌ๋‹จ ๊ธˆ๊ณ ์—์„œ 3์ฒœ๋งŒ ๋‹ฌ๋Ÿฌ๋ฅผ ๋–ผ์–ด ์ƒˆ ์ œํ’ˆ์„ ๋งŒ๋“ค๊ฒ ๋‹ค๊ณ  ๋ฐœํ‘œํ–ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ๋กœ ๋‚ด๋†“์„ ์ œํ’ˆ์€ AI ์ถ”๋ก  ์‹œ์žฅ์ด๊ณ  12์›”์— ์—ด ๊ณ„ํš์ด๋ผ๊ณ  ํ•œ๋‹ค. ์ถ”๋ก ์€ ์ด๋ฏธ ํ•™์Šต๋œ AI ๋ชจ๋ธ์„ ์‹ค์ œ๋กœ ๋Œ๋ ค ๋‹ต์„ ๋ฐ›์•„ ๋‚ด๋Š” ๊ณผ์ •์„ ๋งํ•œ๋‹ค. MEGA๋Š” ๋น„์Šทํ•œ ๋ชจ๋ธ์„ ๋‘๊ณ ๋„ ์—…์ฒด๋งˆ๋‹ค ๋ฐ›๋Š” ๊ฐ€๊ฒฉ ์ฐจ์ด๊ฐ€ ํฌ๊ณ  ์‚ฌ๋Š” ์ชฝ์€ ์ฒ˜์Œ ๋ณด๋Š” ์‹ผ ์—…์ฒด๋ฅผ ๋ฏฟ๊ธฐ ์–ด๋ ต๋‹ค๋Š” ์ ์„ ๋ฌธ์ œ๋กœ ๊ผฝ์•˜๋Š”๋ฐ ์—ฌ๋Ÿฌ ์—…์ฒด์˜ ์ถ”๋ก ์„ ์„œ๋กœ ๋ฐ”๊ฟ” ์“ธ ์ˆ˜ ์žˆ๊ฒŒ ๋งŒ๋“ค์–ด ๊ฒฝ์Ÿ์„ ๋ถ™์ด๊ฒ ๋‹ค๋Š” ๊ตฌ์ƒ์ด๋‹ค. ์ƒˆ ์ œํ’ˆ์€ ์ „๋ถ€ ๋ฉ”๊ฐ€์ด๋” ์œ„์—์„œ ๋Œ์•„๊ฐ€๊ณ  ์˜ˆ์ •๋œ ๋น ๋ฅธ ์™„๊ฒฐ์„ฑ ์—…๊ทธ๋ ˆ์ด๋“œ๋„ ์—ฌ๊ธฐ์— ๋งž์ถฐ์ง„๋‹ค๊ณ  ํ–ˆ๋‹ค. ์–ด๋–ค ๋ชจ๋ธ๊ณผ ๊ณต๊ธ‰์ž๊ฐ€ ๋“ค์–ด์˜ค๋Š”์ง€, ์‹œ์žฅ์ด ์–ด๋–ป๊ฒŒ ๊ตด๋Ÿฌ๊ฐ€๋Š”์ง€๋Š” 12์›” ์ถœ์‹œ ๋•Œ๊นŒ์ง€ ์ผ๋ถ€๋Ÿฌ ๊ณต๊ฐœํ•˜์ง€ ์•Š๊ฒ ๋‹ค๊ณ  The Block์— ๋ฐํ˜”๋‹ค. ํ† ํฐ ์“ฐ์ž„์ƒˆ๋ฅผ ๋ฌป๋Š” ์งˆ๋ฌธ์—๋„ ์“ธ ์ˆ˜ ์žˆ๋Š” ๊ณณ์ด ์žˆ์œผ๋ฉด ์–ด๋””๋“  ์“ฐ๊ฒ ๋‹ค๋Š” ๋‹ต ์ •๋„๋งŒ ๋‚˜์™”๋‹ค. ์ด ๋ฐœํ‘œ๊ฐ€ ๋‚˜์˜จ ๋ฐฐ๊ฒฝ๋„ ๊ฐ™์ด ๋ด์•ผ ํ•œ๋‹ค. The Block์ด ์ธ์šฉํ•œ DefiLlama ๊ธฐ์ค€์œผ๋กœ ๋ฉ”๊ฐ€์ด๋”์˜ ์˜ˆ์น˜ ์ž์‚ฐ์€ ํ•œ๋•Œ 2์–ต ๋‹ฌ๋Ÿฌ๋ฅผ ํ›Œ์ฉ ๋„˜์—ˆ๋‹ค๊ฐ€ ์ง€๊ธˆ์€ ์ฒœ๋งŒ ๋‹ฌ๋Ÿฌ๋Œ€๋กœ ์ค„์—ˆ๊ณ  MEGA ํ† ํฐ์€ ๊ณ ์  ๋Œ€๋น„ 70% ๋„˜๊ฒŒ ๋น ์ง„ ์ƒํƒœ์ด๋‹ค. ์ฒด์ธ ์ž์ฒด๋กœ๋Š” ์ˆ˜์š”๋ฅผ ๋ถ™์žก์ง€ ๋ชปํ•œ ์ƒํ™ฉ์—์„œ ์ง์ ‘ ์ˆ˜์š”๋ฅผ ๋งŒ๋“ค์–ด ๋ณด๊ฒ ๋‹ค๋Š” ์„ ํƒ์œผ๋กœ ์ฝํžŒ๋‹ค. ๋ฐฉํ–ฅ ์ž์ฒด๊ฐ€ ํ‹€๋ ธ๋‹ค๊ณ  ๋‹จ์ •ํ•˜๊ธฐ๋Š” ์–ด๋ ต๋‹ค. ๊ทธ๋Ÿฐ๋ฐ ์ถ”๋ก  ์‹œ์žฅ์€ ์ด๋ฏธ ์Ÿ์Ÿํ•œ ์—…์ฒด๊ฐ€ ๋งŽ์€ ๊ณณ์ด๊ณ  ํ† ํฐ์ด ๊ฑฐ๊ธฐ์„œ ์–ด๋–ค ์—ญํ• ์„ ๋งก๋Š”์ง€๋Š” ์•„์ง ๋นˆ์นธ์ด๋ผ 12์›”์— ๋‚˜์˜ฌ ์‹ค์ œ ๊ตฌ์กฐ๋ฅผ ๋ณด๊ณ  ํŒ๋‹จํ•ด๋„ ๋Šฆ์ง€ ์•Š์„ ๊ฒƒ ๊ฐ™๋‹ค.
3
50