OpenAI's 'Patch the Planet' hands maintainers AI-found CVEs. Who triages them?
OpenAI's 'Patch the Planet' sends AI-found CVEs to OSS maintainers with Trail of Bits review. The whole value lives in the false-positive rate nobody's published.
OpenAI rolled out "Patch the Planet" under its Daybreak security program, with Trail of Bits on the expert-review side. The pitch: AI finds, validates, and fixes vulnerabilities in open-source code, with humans in the loop.
If you maintain a repo, the question isn't whether the model can find bugs. It can. The question is who eats the triage cost.
I've run enough automated scanners to know the failure mode. The tool flags 40 "vulnerabilities," 6 are real, 20 are theoretical paths that require an attacker to already own your box, and 14 are flat-out wrong because the model didn't understand a sanitizer three call-frames up. Guess who reads all 40? The unpaid maintainer.
The interesting detail per OpenAI's announcement is the validation step before anything reaches a human. They claim the AI confirms exploitability rather than just pattern-matching on a suspicious function. That's the whole ballgame. A flag that says "reachable from this public endpoint with this payload" is worth triaging. A flag that says "strcpy here looks scary" is noise you already auto-ignored.
The Trail of Bits piece is what makes me cautiously not-cynical. Expert review by an actual security firm means someone competent stands between the model's output and the maintainer's inbox. If that loop is real — if a human at Trail of Bits validates before the PR lands — this could actually reduce load instead of adding to it. If it's theater, a rubber stamp on model output, it's just a fancier scanner dumping work on volunteers.
What I'd want to see before trusting it: the false-positive rate on real projects, who owns the disclosure timeline, and whether maintainers can opt out of being a free QA queue. None of that is clear from the original report yet.
Verdict: the architecture is right — validate before you escalate, put experts in the loop. But the value lives entirely in the false-positive numbers nobody has published. I'd let it run against a project I maintain in read-only/suggest mode for a month before I'd let it open PRs. Promising, unproven, and the kind of thing that's either genuinely useful or a maintainer's new spam folder. No way to know until the metrics show up.