building @ZeroLeaksAI. prev @orchid_hq
Spain
Joined April 2023
- Tweets1.7K
- Following527
- Followers10.1K
- Likes63.2K
Pinned Tweet
I've just ran @OpenClaw (formerly Clawdbot) through ZeroLeaks.
It scored 2/100. 84% extraction rate. 91% of injection attacks succeeded. System prompt got leaked on turn 1.
This means if you're using Clawdbot, anyone interacting with your agent can access and manipulate your full system prompt, internal tool configurations, memory files... everything you put in SOUL.md, AGENTS.md, your skills, all of it is accessible and at risk of prompt injection.
For agents handling sensitive workflows or private data, this is a real problem.
cc @steipete
Full analysis: zeroleaks.ai/reports/opencla…
Lucas Valbuena retweeted
Introducing Shield 1.1
Shield is the most accurate detector for prompt injections and jailbreaks. AI agents read text they didn't write: web pages, tool results, MCP tool descriptions, skill files.
Any of it can carry instructions that hijack the agent. Your app sends that text to Shield before the agent acts on it, and Shield tells you whether it's an attack. One API call.
Shield 1.1 makes Shield Large, our most accurate model, better.
Lucas Valbuena retweeted
Introducing Shield 1.1.
Shield checks the text your agent reads (tool results, web pages, MCP tool descriptions, skill files) for prompt injections and jailbreaks before the agent acts on it.
In 1.1, Shield Large catches more attacks and flags fewer benign texts: 4.1%, down from 5.4%, and lower on all six false-alarm datasets. Prompt templates went from 7.0% to 2.3%, coding-agent sessions from 5.3% to 2.2%.
zeroleaks.ai/blog/shield-1-1
teapawt.org/lucknite/status/211585…
Introducing Shield 1.1
Shield is the most accurate detector for prompt injections and jailbreaks. AI agents read text they didn't write: web pages, tool results, MCP tool descriptions, skill files.
Any of it can carry instructions that hijack the agent. Your app sends that text to Shield before the agent acts on it, and Shield tells you whether it's an attack. One API call.
Shield 1.1 makes Shield Large, our most accurate model, better.
Introducing Shield 1.1
Shield is the most accurate detector for prompt injections and jailbreaks. AI agents read text they didn't write: web pages, tool results, MCP tool descriptions, skill files.
Any of it can carry instructions that hijack the agent. Your app sends that text to Shield before the agent acts on it, and Shield tells you whether it's an attack. One API call.
Shield 1.1 makes Shield Large, our most accurate model, better.
We ran 14 other detectors on the same 43 sets with the same scorer, each at its default setting: paid guardrail services, open models, and GPT-6.1 Sol as an LLM judge.
Shield 1.1 is first of 16 overall at 91.5. The GPT-6.1 Sol judge scores 89.0, and it needs a language-model call for every text. Jev by TypeSafe, the best of the rest, scores 84.6.
ALT Chart: overall score of 16 detectors on the same benchmark. Shield 1.1 91.5, Shield 1 90.4, GPT-6.1 Sol as a judge 89.0, Jev by TypeSafe 84.6, OpenAI Guardrails 80.1, Patronus Wolf Defender 79.6, Horizon-Labs guard large 78.7, Amazon Bedrock Guardrails 72.8, TestSavant defender large 71.0, Sheltron 68M 68.3, PIGuard 66.3, Azure AI Prompt Shields 63.2, ProtectAI v2 (LLM Guard) 61.0, Meta Prompt Guard 2 86M 58.6, deepset DeBERTa injection 57.2, NVIDIA NemoGuard 49.0 (fewest false alarms, but flags 2% of attacks).
Model card: zeroleaks.ai/shield/shield-1…
Technical report: zeroleaks.ai/shield/shield-1…
Lucas Valbuena retweeted
Every landing page has the same flat background.
So I built gradients.sh, moving gradient backgrounds for any site.
- 207 animated presets
- Tweak colors, speed and grain
- React, Vue, Svelte or plain HTML
- Or ask your agent to add it
39 free, all customizable 👀
Lucas Valbuena retweeted
the question we got asked the most is whether our viral launch video was ai-generated
it wasn’t. we risked our lives for pancake.ai
why? watch to find out
Meet Pancake, the Vercel for GTM.
Try it at pancake.ai
Your agent ships your product. Now it ships your go-to-market.
How? A GTM brain that knows your business, 50+ data providers to find the right customers and tools to reach them.
Your next customers are one prompt away.
(did I really have to jump off a plane to say that? questionable, but I survived and now you're reading my tweet)
Lucas Valbuena retweeted
if you sell AI training data, see how you compare: we found 159 companies selling data, RL envs and evals, every fact sourced
plus 38 programs that pay companies for their data: whosellsdata.com
8 of the 14 that pay companies directly publish a dollar range. most of the rest set terms per deal. unknowns are marked and nothing is ranked
the other page: what each program wants, who qualifies and what it offers: whosellsdata.com/buying
missing something, or something wrong? fix it on the site, a person reviews every change
we built this at @Secludy. if you buy, broker or license company data and need the names out before it reaches a lab, dm me
Lucas Valbuena retweeted
codex limits are VERY bad. i’ve used both of my 200 Pro plan accounts within a day 🙃
