Trail of Bits may end up inside OpenAI, report says

OpenAI is in early talks to buy Trail of Bits, the security firm that finds and fixes vulnerabilities in open-source software and already runs a bug-fixing program with OpenAI, The Information reported on October 13. So far it is one report resting on unnamed sources, and those sources said the talks could still fall apart. The size of the deal is not known.
The Information says the acquisition would help OpenAI expand its security business. That business has been a contest this year: OpenAI and Anthropic have both built specialised AI for finding software vulnerabilities. OpenAI has also just hired Dally Rajic, a former executive at Wiz, the Google-owned cloud security company, as chief revenue officer.
Trail of Bits was founded in 2012 by Dan Guido, who is still its CEO. A 2026 conference bio puts its staff at more than 140 engineers, with expertise in cryptography, AI and blockchain security. Besides audits, the firm has helped maintain open-source projects.
OpenAI announced Patch the Planet, its joint program with Trail of Bits, on June 22. Trail of Bits engineers use Codex and security-focused models on projects such as Python, and they review what the tools find before it reaches maintainers. They then work with each project on patches and tests, according to TechCrunch, which said at launch that it was hard not to read the program as a competitive swipe at Anthropic. TechCrunch also said it was unclear how the program would work in the long term, or whether it would scale up at all.
The program’s output is already public. A Trail of Bits audit of OpenClaw under Patch the Planet turned up a bug that let agents keep reading memory after memory access was turned off, and the auditors checked every finding by hand before reporting it.
Guido credits the tooling with the pace. “Thanks to Codex, we were able to deliver in a day what would have taken weeks,” he said.
Up next
SubQ Snap reads 384K tokens before each agent decision
384K tokens per agent decision is 12 times Jev’s 32K limit, and Subquadratic calls even that a preview cap, since SubQ Snap was tested at one million tokens. Who picks your agent’s next tool now, a small model or the main LLM?
Related stories
- OpenAI to buy Pierre, whose diff tools already run in Codex
- OpenAI models helped find critical flaws in PHP
- OpenClaw agents kept reading memory after it was disabled
- Hugging Face wants OpenAI to pay the breach bill in GPUs
- $1B in subsidized models for water and power utilities
- Preparedness team gone, risk calls split by subject
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

