🌊🐝 BD @trailofbits
- Tweets 1,547
- Following 4,132
- Followers 469
- Likes 4,148
Colby retweeted
We're the only red team for @ARIA_research's Safeguarded AI programme, led by @ammannnora.
Each development cycle, we attack the code, its proofs, and everything those proofs rest on: the models, specs, assumptions, and deployment choices behind every team's assurance case.
AI-enabled cyber attacks are getting faster + going further. But what if AI-enabled formal methods could make high-assurance cyber defence practical at unprecedented speed + scale? We’re funding eight teams with £22m to test that idea. Meet them here: link.aria.org.uk/SAI-TA2-Cre…
3
8
47
6,253
Colby retweeted
We're hiring a Senior Developer Relations Engineer who builds and ships, with the security depth to work alongside our engineers and open-source maintainers.
Apply here: apply.workable.com/trailofbi…
1
3
20
2,613
Colby retweeted
We reviewed @bee__computer Private Compute's mobile apps, servers, and AMI builder and reported 21 issues. Amazon fixed all 7 high-severity findings and published the full report.
github.com/trailofbits/publi…
Personal AI should have frontier intelligence without sacrificing privacy. That's why we built Bee Private Compute.
Bee believes trust is built on transparency, so today we are publishing our whitepaper and a third-party audit performed by @trailofbits.
3
3
53
7,021
The biggest reduction in false positives we’ve seen in AI auditing came from giving the agent a clear threat model.
Put a THREAT_MODEL.md at the repo root and tell the agents to read it.
Cover:
• What the system does and how it’s deployed
• What needs protecting
• Who is trusted, and for which actions
• What an attacker can access or control
• Where untrusted input enters
• Which security rules must always hold
• What’s out of scope, and why
With the latest SOTA models, we’re seeing fewer hallucinations but more findings based on incorrect assumptions about the system, which is now the main source of false positives.
Codex Security’s threat-model skill is a good starting point, but the devs should own the document, verify every claim, and keep it up to date as the codebase evolves.
github.com/openai/codex-secu…
12
44
317
15,448
Colby retweeted
We’re hiring an Engineering Director to lead Application Security. We want someone who still finds bugs in code, builds excellent teams, and has a clear view of how AI impacts security audits.
Apply here: apply.workable.com/trailofbi…
2
8
53
4,002
Colby retweeted
today we announced OPEN-1B, the world's first fully auditable transformer training run
OPEN-1B proves that we can log and audit every single step of AI training and inference, providing proof of its training data, recipe, biases, and weights
open1b.gensyn.ai/
10
32
140
24,082
.@trailofbits released 3 skills to help write patches:
`post-patch-validation` helps agents test that fixes are correct
.@1Password's FLAWED report says AI models produce a clean security fix only 26% of the time.
Defenders shouldn't take that number seriously.
• The six vulnerabilities were handpicked because their fixes were complex. Clean-fix rates ran from 3% to 60% depending on the bug, and the report averaged them together.
• Agents set up to fail were counted in the headline figure. Two of 1Password's prompts instructed the agent to apply the wrong fix. Those trials make up 22% of the data. One evaluation mode prevented the agent from compiling or running any code, and it accounts for 36% of the data.
• The report ran two models, GPT-5.5 at medium effort and Opus 4.8 at high. Neither was tested at its highest available setting, so the report says nothing about how more effort or stronger models change the results.
• Several instruction and grading errors further undercut the headline, and are elaborated upon in the attached blog.
We've spent four months submitting hundreds of AI-authored patches to widely adopted open-source projects as part of Patch the Planet.
Our experience didn't match 1Password's report, so we did a full analysis across 186 AI-authored pull requests and 33,500 subsequent commits, benchmarked against 2,265 human-authored patches we graded across years of security engagements. blog.trailofbits.com/2026/09…
11
17
119
216,741
Personally, I think a team like @trailofbits would have been a far better selection for a qualified, independent third-party to perform these sorts of reviews. 🤷♂️
25
26
280
39,019
Colby retweeted
Our engineer Joe Doyle built trailmix, the quantum circuit toolkit we open-sourced in June. The paper's leading circuits build on it, and his leaderboard submissions helped push the score past Google's. github.com/trailofbits/trail…
The first full paper on ECDSA.fail is on arXiv
In March, Google Quantum AI reported a more efficient quantum circuit for a core step in breaking the signatures behind Bitcoin and Ethereum. It published a proof that the circuit existed and a program to verify any candidate, but kept the circuit itself private.
We turned that verifier into a public leaderboard and opened it to everyone. Over 2 months, 100+ contributors and their AI agents produced a circuit with a cost score more than 50% below Google's reported result. The live leaderboard has since moved to 62% ahead.
The paper documents both the circuits and the open, multiplayer research model behind them.
That model is now @YukonResearch.
And the challenge is still open.
ALT From a Hidden Quantum Circuit to Open Autoresearch: The ECDSA.fail Story
4
13
76
7,365
Colby retweeted
A short and truly marvelous proof of Fermat's Last Theorem? Not quite. But with GPT, I did find a delightfully on-theme Lean bug that lets a 20-line proof fit in the margin. Thankfully, not a kernel soundness bug (read the full blog post!), and it was quickly fixed!
Last week, @AnthropicAI formalized Fermat's Last Theorem in 13 million lines of Lean code.
We "proved" that same theorem in 20 lines by exploiting a bug we found in Lean. blog.trailofbits.com/2026/09…
2
15
569
Colby retweeted
My teammate “proved” Fermat’s Last Theorem and did something Fermat or anyone else couldn’t do: made it small enough to fit in the margin
Last week, @AnthropicAI formalized Fermat's Last Theorem in 13 million lines of Lean code.
We "proved" that same theorem in 20 lines by exploiting a bug we found in Lean. blog.trailofbits.com/2026/09…
1
2
24
1,414
Colby retweeted
OpenSSL's latest advisory covers 9 patched vulnerabilities. Our engineers Filipe Casal and Opal Wright found 4 of them as part of Patch the Planet.
openssl-library.org/news/sec…
14
44
4,547
Colby retweeted
aiohttp is downloaded 600M+ times a month. Its maintainers just recapped a year of security work, including the 10 issues (8 CVEs) we reported through Patch the Planet.
Every fix shipped days later in 3.14.1. aio-libs.org/news/2026/secur…
3
4
26
3,205
Colby retweeted
1,535 potential bugs found, 1,017 awaiting patches, 326 fixes open upstream, 192 merged.
Since Aug 4: +398 bugs, +46 merged, 55 codebases under review, and scapy has overtaken vllm as the most-reported codebase in Patch the Planet.
trailofbits.com/patch-the-pl…
1
12
37
4,699
Colby retweeted
We launched as a trusted auditor for Signal's Automatic Key Verification, watched GPT 5.6-Cyber chain three 0-days into a VM escape, and found a Provenance bug representing ~$500K in drainable HASH. Plus 8 new public reviews, 2 open-source tools, and more.
August Tribune: mailchi.mp/trailofbits/augus…
10
11
111
12,367
Colby retweeted
We signed the call for collective action. Across 3 months, we audited 50 open-source projects using frontier cyber models. To date, we've found 1,268 potential issues and wrote 298 patches.
Critical infrastructure around the world carries the same technical debt, but has far fewer defenders. It will take the entire industry working together to fix it.
An open letter for a global surge in cyber defense, signed by over 100 organizations including Anthropic, AWS, Google, Microsoft, OpenAI, and Oracle.
13
8
108
12,929
AI is calling in all of our security debt, all at the same time.
Most sandboxes were never very secure, but that didn't matter because expertise was scarce and effort was expensive. Not so in 2026.
We asked GPT 5.6-Cyber to escape a VM used to sandbox agents. It broke out three times.
In its final escape, the agent found three 0-days on its own and chained them into a working exploit. blog.trailofbits.com/2026/08…
8
44
205
22,919
Colby retweeted
We asked GPT 5.6-Cyber to escape a VM used to sandbox agents. It broke out three times.
In its final escape, the agent found three 0-days on its own and chained them into a working exploit. blog.trailofbits.com/2026/08…
70
333
1,696
491,590
Colby retweeted
We found a bug that lets any user grant themselves admin access to live financial assets on Provenance, a Cosmos SDK chain.
82 token accounts were exploitable, representing ~$500K in drainable HASH. Now patched. blog.trailofbits.com/2026/08…
2
10
144
16,512
Colby retweeted
PATCH THE PLANET BUG SPOTLIGHT: We found a medium-severity bug in aiohttp, Python's HTTP engine that had 600M+ downloads last month. Denys Pakizh caught oversized requests dodging its size limits. Now patched. CVE-2026-54277 in the dashboard: trailofbits.com/patch-the-pl…
3
7
36
5,131