AI coding agents accelerate security exploit discovery within minutes of bug disclosure

A Cambridge computer science professor and OCaml compiler maintainer reports that automated probes for security vulnerabilities begin within minutes of patches being shared for public discussion. The maintainer observed attempts to exploit percent-encoded traversal sequences approximately ten minutes after a patch was posted, indicating that automated systems are monitoring public repositories for new disclosures.

The maintainer demonstrated that modern coding agents can turn minimal hints about a bug into actionable exploit paths, switching from one commercial model to another after the first declined to proceed with the task. This suggests that the threshold for turning a theoretical flaw into a practical exploit has dropped significantly due to AI-assisted tooling.

The maintainer argues that this pace of discovery is incompatible with existing open-source embargo practices, which typically allow maintainers days or weeks to prepare and release fixes before public disclosure. The rapid automation of exploit discovery may require new processes to keep communities safe while preserving transparency.

The maintainer of the rclone project independently corroborated the trend, noting a sharp increase in security disclosures. Over the first ten years of the project, the maintainer received about 20 security disclosures via GitHub, but in the preceding month alone, the project handled more than 40 disclosures. The maintainer described using AI tools to triage and draft fixes, achieving a roughly 75% hit rate in identifying issues that required attention.

The surge in disclosures has also strained the CVE assignment process. Before what the maintainer termed the “AI apocalypse,” GitHub typically assigned CVEs within two to three days, but now assignments are taking three to four weeks. As a result, the maintainer has resorted to shipping point releases with placeholder “CVE-PENDING” labels in changelogs, a practice that complicates downstream security tracking.