AI agents turn CVE descriptions into working exploits within minutes

Anil Madhavapeddy, a computer science professor at the University of Cambridge and a core maintainer of the OCaml compiler, has written that AI agents can turn public clues about software flaws into working exploits, weakening the traditional disclosure embargo used by open source projects. He described his own experience fixing a path traversal vulnerability. The patch itself was not complicated, and the usual security process would have been to fix it privately, notify affected users, and then publish a public advisory. This time, however, just minutes after he submitted the pull request for the fix, the web server logs showed probing requests that exactly matched the pattern of that vulnerability.

He cited a study in which a GPT-4 agent successfully exploited 87% of the vulnerabilities in a benchmark containing 15 flaws when it was given the CVE description, compared with only 7% when no CVE description was available. Madhavapeddy argues that the current "economics of vulnerabilities" now work against open source maintainers, adding that if even one person is searching for a certain class of flaw, it may show up as a question on a mailing list or activity on an isolated branch.

He therefore argues that security processes need to be inverted to some degree, shifting away from secrecy toward faster patch releases and protocol-level mitigations, and that patch development and release cycles must speed up further. The matter currently rests at the publication of his article and these arguments; the material does not mention any responses from other parties or any follow-up actions.