On 28 August 2026, Simon Willison relayed a report from OCaml core maintainer Anil Madhavapeddy: within ten minutes of a patch being shared for discussion, his site was fielding probes for percent-encoded traversal sequences. This is a striking acceleration from the days-to-weeks window that open-source embargo practices were designed around.1

Context

Madhavapeddy attributes the speed to coding agents that can turn a hint of a bug into an exploit, and says he demonstrated the effect with his own agents, switching to DeepSeek V4 Pro when Claude Fable refused the task. rclone maintainer Nick Craig-Wood reports a parallel pressure: over 40 security disclosures in the last month, versus about 20 in the project's first ten years.1

Assessment

The significance is not that the internet scans for vulnerabilities, which is routine, but that a patch discussion itself now appears sufficient signal for targeted, exploit-like probes. If representative, this invalidates the assumption that a few days of embargo after public patch discussion is safe; the binding constraint moves to how quickly maintainers can triage AI-amplified disclosure volume while CVE assignment lags by weeks.1

Caveats

The evidence is a small set of anecdotes from large, actively scrutinized projects; systematic data on disclosure rates across the ecosystem, and on what fraction of probes are malicious versus automated scanning noise, are absent. rclone's numbers conflate submissions with validated vulnerabilities, and the 75% 'nugget' rate does not mean 75% were exploitable. Systematic data showing these probes are ordinary scanner noise, or that disclosure volume is confined to high-profile projects, would undercut the urgency.1