AI agents now exploit a public bug-fix PR within minutes of opening
Maintainers report probes seconds after a fix goes public and a 40-in-a-month disclosure surge; GitHub's CVE turnaround has slipped from days to weeks.
Copy markdown
Ten minutes from PR to probe
Cambridge's Anil Madhavapeddy opened a public fix for an OCaml path-traversal bug and within about ten minutes his server logs showed automated agents firing percent-encoded traversal probes at the exact flaw — before any release shipped. The private embargo window builders rely on is effectively gone.
A rumor of a bug is all it takes
Attackers don't need your patch, just a hint. Madhavapeddy rebuilt a working exploit in under a minute with DeepSeek V4 Pro after Claude Fable refused, and cites research showing LLM agents exploit 87% of vulnerabilities handed a CVE description. He now pegs mean-time-to-exploit at roughly -7 days: attacks land before the fix does.
rclone: 40 disclosures in a month, up from ~20 in a decade
rclone maintainer Nick Craig-Wood says the project fielded 40+ security disclosures last month versus about 20 across its entire first decade, and roughly 75% are real issues needing triage. Solo and small-team maintainers are the ones drowning in the new volume.
CVEs now take weeks, so fixes ship 'CVE-PENDING'
GitHub's CVE assignment has stretched from 2-3 days to 3-4 weeks, pushing maintainers to ship patches tagged 'CVE-PENDING.' If your dependency scanners key on CVE numbers, a fixed hole in your stack may go unflagged until the identifier finally lands.
What this changes for how you ship
The advice is shifting from secrecy to throughput: develop fixes in private infrastructure with vetted reviewers, batch and ship security releases continuously instead of telegraphing them in public PRs, and consider protocol-layer 'virtual patches' you can deploy before the upstream fix is final.