A rumour of a bug is now enough for AI to build the exploit

The old bargain of coordinated disclosure was that keeping a vulnerability quiet until a patch shipped gave defenders a head start. Anil Madhavapeddy, a systems researcher at Cambridge, argues that this no longer holds. An AI agent now needs only a rough hint about where a bug lives to find and weaponize it, often faster than the maintainer can push a fix. He tested this himself: given the cohttp path traversal bug, he had DeepSeek V4 Pro produce a working exploit in under a minute. Ten minutes after he opened the public pull request with the fix, automated probes for that exact flaw showed up in his server logs.

The pattern is not just his. He points to work by Fang and colleagues showing that agents exploited 87% of vulnerabilities when handed the CVE description, but only 7% without it, so the written description is most of the weapon. He cites recent cases where the gap from disclosure to exploitation collapsed to hours: nine hours for a marimo CVE, twenty for one in Langflow, neither of which had public proof-of-concept code to copy. By his reading the "mean time to exploit" went negative, meaning working attacks now tend to appear before the patch reaches most users.

His point is not to give up on disclosure but to change what defenders build around it. He suggests better discussion infrastructure for open source projects, a habit of shipping fixes continuously rather than in slow batches, stronger cross-ecosystem package management so patches propagate quickly, and defenses at the protocol layer that can be deployed faster than an upstream fix can travel.

Why it matters

If you maintain open source software, treat the moment your fix PR becomes public as the moment exploitation starts, not a safe window before it. The practical response is faster propagation of patches and defenses you control, because the description of a bug is now most of the work an attacker needs.