For most of 2026, OpenAI has been running an unusual experiment: pointing its own AI models at the open-source software that underpins much of the internet, to see how many security flaws they can find — and, harder still, how many they can actually get fixed. The early results, visible on a public dashboard the company updates with partners, show a company that has gotten very good at the first part and is still struggling with the second.
The project, called Patch the Planet, is built with the security firm Trail of Bits and launched in June as part of OpenAI's broader "Daybreak" cybersecurity push. As of its most recent public snapshot, the dashboard listed 68 open-source codebases under review, 1,995 issues found, 679 patches written by the AI system — and only 466 of those accepted upstream by the maintainers who actually run the projects. That is roughly a quarter of everything flagged.
Running alongside it is Aardvark, an autonomous agent OpenAI pairs with its Codex coding tool. Aardvark reads a codebase the way a human security researcher would, reasons about how a change might introduce a vulnerability, and then proposes a validated patch rather than just a bug report. In OpenAI's own benchmark against "golden" repositories seeded with known and artificially introduced flaws, the system identified 92 percent of them — and in real-world use, at least ten of its findings have been serious enough to be assigned formal CVE identifiers, the public numbering system used to track vulnerabilities industry-wide.
The patching gap
None of this is happening in a vacuum. Security researchers have already shown that general-purpose AI models can find real, previously unknown flaws — a researcher's use of OpenAI's o3 model last year surfaced a genuine zero-day in the Linux kernel's networking code, which was patched and merged into the mainline kernel. What's new is the scale: instead of one researcher with one model, OpenAI is now running a standing pipeline across dozens of projects at once, with HackerOne handling intake, triage and coordinated disclosure, and participating maintainers getting a week of Trail of Bits engineering time plus six months of ChatGPT Pro access in return for letting the AI loose on their code.
The company argues that is the point. "AI is accelerating vulnerability discovery, but discovery alone does not protect users," OpenAI said of the program. The harder problem, in other words, isn't finding bugs anymore — it's getting the people who maintain software, many of them unpaid volunteers, to review, trust and merge fixes as fast as an AI can generate them. Wharton professor Ethan Mollick, who tracks enterprise AI adoption, has noted that companies he talks to are already struggling with the sheer rate of patches AI tools are producing, a strain that shows up starkly in Patch the Planet's own numbers: more issues found than patches written, and more patches written than maintainers have had time to accept.
OpenAI is not alone in running this kind of audit. Anthropic conducted a parallel scan of open-source software the same week Patch the Planet's latest figures were published, and it showed the same shape — a large pool of candidate findings next to a much smaller number that made it into a shipped release. The pattern suggests the bottleneck isn't particular to one company's tooling; it's a structural mismatch between how fast frontier AI can now generate security work and how fast the human infrastructure around open-source software can absorb it.
What happens next
For now, Patch the Planet and Aardvark remain limited in scope: a few dozen codebases, a private beta, and partner maintainers who opted in. But the direction of travel is clear enough that it is reshaping how security researchers think about the next few years — not as a question of whether AI can find dangerous bugs in the software billions of people rely on, but of whether the unpaid, often part-time maintainers of that software can keep up with what AI finds. HackerOne's coordinated-disclosure role is meant to prevent findings from leaking before fixes exist, but as more labs run similar programs, the volume moving through that pipeline is only likely to grow. Whether open-source maintenance gets more funding, more automation, or simply falls further behind is, by OpenAI's own numbers, still very much an open question.
