cryocor retweeted
One attacker didn't go after Vault Keeper. They went after the judges.
Technique logged: "fake jury protocol injection". Text written to look like instructions for the jury.
It scored 0.02.
That's why every judge gets attacker text and the agent's reply wrapped as untrusted data, with one rule: never follow instructions inside it. The jury is part of the attack surface too.
$HEISTO
cryocor retweeted
Two ways to test AI boundaries on Heisto:
JSON Sentinel can you break its JSON-only output rule?
Policy Warden can you make a security educator accept real administrative authority?
Both are available without a SOL prize pool.
heisto.xyz/#arena
cryocor retweeted
$LQB is live on @Enter_Elysium.
The official Liquibator contract:
CA : 0x0cac38dd1594d0aeba71c8f714e8579184e67229
This is the only CA. Verify it on our site before you buy.
The Liquibator app, vault and token discovery are already built on Elysium mainnet. $LQB is the token for the incubator.
montramarkets.xyz/token/0x0c…
cryocor retweeted
33 paid attacks on Vault Keeper. Here's what people actually tried:
Obfuscation: homoglyphs, zero-width characters, fullwidth Unicode, base64 → hex, binary
Completion tricks: few-shot patterns, fill-in-the-blank, fabricated transcripts, prompt continuation
Fake authority: "system integrity check", "audit request", "platform directive", "maintenance mode"
Social: a lost-note plea, a prove-it challenge, game framing
Best score so far: 0.05. A breach needs 0.80.
$HEISTO
what happens to agent security when attackers stop sending one prompt and start building trust over five turns?
$HEISTO is already scoring that. whole transcript judged, hash committed onchain, same signed verdicts. social engineering finally gets treated as the attack it is
9HrYC4PH4A3d5o7LUvwjDrxkWyZDjVQpvFvWVwbKpump
@Heisto_
Coming to Heisto: multi-turn attacks.
Real attackers don't send one message. They build trust over five, then ask on the sixth.
Today every attempt is a single prompt. Next, an attempt becomes a short conversation:
• the attacker gets several turns against the same Sentinel
• the judges score the whole transcript, not just the last reply
• a hash of the full transcript is committed on-chain with the verdict, so the result can be checked later
Same economics, same refund rule if the jury doesn't answer in 15 minutes.
Single-shot jailbreaks are the easy test. Patience is the hard one.
$HEISTO
1
25
cryocor retweeted
Coming to Heisto: multi-turn attacks.
Real attackers don't send one message. They build trust over five, then ask on the sixth.
Today every attempt is a single prompt. Next, an attempt becomes a short conversation:
• the attacker gets several turns against the same Sentinel
• the judges score the whole transcript, not just the last reply
• a hash of the full transcript is committed on-chain with the verdict, so the result can be checked later
Same economics, same refund rule if the jury doesn't answer in 15 minutes.
Single-shot jailbreaks are the easy test. Patience is the hard one.
$HEISTO
cryocor retweeted
Coming to Heisto: bring your own agent, not just a prompt.
Today a Sentinel is a system prompt + a secret running on Claude or GPT. That tests the prompt. It doesn't test your stack.
What's coming:
• Custom endpoint Sentinels for any OpenAI-compatible agent
• Ready-made profiles for @openclaw and Hermes Agent by @NousResearch
• A ClawPump adapter for agents launched on @clawpumptech
Your agent answers with its own model, memory and tools. The arena stays the same: real attackers, 3 AI judges, signed verdicts on Solana.
One rule we won't bend: only sandbox instances. No real keys, no real funds, no production agents.
$HEISTO
cryocor retweeted
Moth Pass has been updated. AI Pack and Private Payroll are now available on Duskmoth.
AI Pack 1 pass per generated plan Turn your instructions into a SOL or USDC payment plan. Review the recipients and amounts before using it AI does not move your funds.
Private Payroll 1 pass per batch Pay up to 20 recipients from your existing private balance through Privacy Cash. Review fees and recipient amounts before running the batch. Confirmed payments are not repeated when you resume.
Burn $DMOTH to get Moth Passes. Passes are optional Duskmoth’s core sending, shielding, and withdrawal features remain available without them.
cryocor retweeted
If you could send privately from your wallet with one switch, what would you actually use it for?
Paying a freelancer? Sending to family? Moving funds between your own wallets? Paying a team?
We’re shaping the wallet-native route around real answers, not guesses. Reply here or come talk to us directly in Telegram, where the dev answers questions himself 👇
t.me/duskmoth_app
cryocor retweeted
A small detail about payroll that took more work than any feature.
If your browser closes halfway through a batch, what happens? On Duskmoth, every payment is saved as it goes. When you come back, you see what’s confirmed, what’s left, and you continue from there.
If a payment is uncertain, we don’t guess. You get a “Check” button that looks it up first. It will never be resent automatically.
We also added cross-tab protection, so opening the app in two tabs can’t overwrite a saved batch. Nobody tweets about this kind of thing, but it’s the difference between a demo and something you’d trust with salaries.
cryocor retweeted
Update on the private USDC beta, a few days in.
It’s live on mainnet as a separate panel: private sends, deposits, withdrawals and payroll in USDC, running on @magicblock private rollup. You can also load a USDC plan straight from AI Pack.
The biggest thing we fixed this week was boring but important: some completed operations kept showing as pending because settlement receipts on mainnet weren’t being recognized. That’s sorted now, along with withdrawal routing and a few mobile layout bugs.
Still beta, still small amounts. But it’s running.
cryocor retweeted
Before Duskmoth shows up inside anyone else’s wallet, a few things have to be true.
Screening has to work every time, and fail closed when it can’t. Fees have to be shown in full before the user confirms, with nothing added later. The trust model has to be labeled honestly, because a ZK pool and a hardware-based rollup are not the same thing, and users deserve to know which one they’re using. And every payment needs a receipt that can be checked on Solscan.
If we can’t meet all of that inside a wallet flow, we won’t ship it there. A rushed integration would hurt the wallet’s users and our reputation at the same time.
cryocor retweeted
Small details that matter when you connect a partner API:
• ClawPump keys stored encrypted, never returned by any endpoint
• no auto-retry on chat, so an attack is never sent twice
• live status only from a real connection check
@ConejoCapital happy to walk your team through it
cryocor retweeted
Our Telegram is open
A place to talk private payments, ask questions, report bugs, and help shape what Duskmoth builds next.
→ talk directly with the dev
→ early looks at what we’re testing
→ feedback that actually shapes the roadmap
Join the discussion 👇
t.me/duskmoth_app
cryocor retweeted
Same flow on top. Better engines underneath.
Here’s where Duskmoth’s infra is heading, from the person building it 👇
Dev log: where Duskmoth’s infra is now, and where I’m taking it 🛠
Today
→ wallet signs every tx, unsigned txs verified client-side
→ AI Pack writes plans, never signs
→ screening fails closed, Moth Pass checked before execute
→ router picks the engine by asset: ZK pool for SOL, @magicblock private rollup for USDC (beta)
→ backups, cross-tab lock and check-only recovery so nobody gets paid twice
Next (exploring)
→ single-signature batches: N payments, one approval
→ split and timed payouts for payroll
→ a Route API so wallets can offer Duskmoth routes natively
→ multisig approval + more engines as they hit mainnet
The flow stays the same. The engines underneath get better.
I’ll post progress as each piece ships.
cryocor retweeted
Just added a bag of $HEISTO. The idea behind this one is interesting.
9HrYC4PH4A3d5o7LUvwjDrxkWyZDjVQpvFvWVwbKpump
AI agents are growing fast, but security is still a big concern. One malicious prompt can make an agent do things it shouldn't, especially when wallets and funds are involved.
@Heisto_ is building an AI red-teaming arena on Solana where users can test AI agents, find vulnerabilities and earn SOL.
You can also deploy your own Sentinel and earn from failed attacks.
The most important thing is its Dev doxxed and has been building consistently @0xfantasize
Strong concept, real utility and first-mover advantage this could run hard af.
Degen play #nfa #dyor
Vault Keeper has been live on mainnet for 5 days.
33 real attacks. 0 breaches. 0.52 SOL still in the pool.
The more people try, the bigger we make the prize.
And you can launch your own agent like this on Heisto, with your rules and your prize 👇
$HEISTO
