Niels Provos: the point of AI bug-finding is proof a maintainer can trust

Finding candidate bugs is not the goal, according to Niels Provos. He is a security researcher who co-authored the bcrypt password-hashing method. He argued that the goal is proving each bug is real. He made the point on GitHub's podcast, in an episode published on 13 November 2026.

What was said

Provos began with open-source maintainers, many of whom, he said, run their projects on the side. He argued that money alone does not fix this. Someone busy with their regular life, he said, cannot do much with funding, because they would have to hire somebody.

Then he turned to AI. Everybody is using it, he said, and maintainers are getting lots of pull requests (proposed code changes) of dubious quality. He recalled hearing that the FFmpeg or curl maintainers were overwhelmed by security reports and said they could not use any of them without patches. He was not sure which project it was.

Provos tied this to the vulnerability discovery workflow he has been working on. In his words, the point is not just finding potential problems. It is to ask, "can you prove that these are really real vulnerabilities?" That requires the model to write a proof-of-concept exploit, a small program that shows the flaw can be used. A maintainer can then know without doubt that the issue is real and spend very limited time on it.

He compared this to security operations teams. "so much of the security industry is about the, we will detect bad things, and you can't detect perfectly," he said. Those teams, he added, spend most of their time on false-positive alerts. He saw open source as very similar.

Why it matters

Our reading: when a tool produces bug reports at volume, the work of checking them lands on someone. If the checkers are unpaid volunteers, the cost lands on them. This is our inference, not Provos's claim. A report with a working exploit moves the checking back to the finder.

For companies that depend on open-source code, this suggests a test for any AI security tool or vendor. Ask whether it shows proof for each finding or only a list of suspicions. For teams that file reports, it suggests attaching evidence, and a patch where possible, as Provos described maintainers asking for.

The other side

Provos described what he is building and what maintainers need. He did not give data on how many reports are false, or show that exploit-backed reports reduce the load. He was also unsure which project's maintainers made the complaint he cited.

Proof also has limits. A model that cannot write a working exploit has not shown that a bug is absent. Provos himself said detection can never be perfect. The conversation also left open who will review and patch the confirmed bugs, since he said maintainers cannot simply buy more hours with money.