AI coding agents turn bug rumours into exploits within minutes, overwhelming open source maintainers

Anil Madhavapeddy, a professor of computer science at Cambridge and core maintainer of the OCaml compiler, has described a new reality where security issues in OCaml projects face attempted exploits within minutes of patches being shared for discussion. In a blog post, Madhavapeddy noted that a website he manages received probes for percent-encoded traversal sequences roughly ten minutes after a discussion began — suggesting automated watchers are monitoring public repositories. He demonstrated that modern coding agents can find flaws with minimal hints, switching from Claude Fable to DeepSeek V4 Pro when the former refused the task.

Nick Craig-Wood, maintainer of the file-syncing tool rclone, confirmed the trend in comments on Hacker News. He stated that in the first ten years of the rclone project, it received about 20 security disclosures through GitHub. "We had to deal with over 40 in the last month!" he wrote, adding that the hit rate for disclosures is about 75%. Craig-Wood noted that GitHub previously assigned CVEs within two to three days, but now takes three to four weeks, forcing him to ship point releases with "CVE-PENDING" in changelogs — an arrangement he described as "not ideal."

The rapid exploitation window appears incompatible with existing open source embargo practices, which assume patches can be discussed privately for several days before a public release. Madhavapeddy argued that the community needs new processes to keep projects safe.

Why This Matters

Open source maintainers, already stretched thin, face a flood of security disclosures that is now accelerating beyond human capacity to triage.
The collapse of traditional embargo timelines means vulnerabilities may be exploited before patches are even ready, threatening the security of millions of downstream users.
If AI agents can routinely turn rumours into exploits, the entire open-source security model — built on coordinated disclosure — may need to be rethought.

Background

Open source security has long relied on a system of responsible disclosure: researchers privately report vulnerabilities, maintainers prepare patches under embargo, and the fix is released simultaneously with public details. This process typically took days to weeks. AI-powered coding agents, capable of reading patch diffs and constructing attacks, have collapsed that timeline to minutes. The problem is not limited to any one language or project — rclone (Go) and OCaml are both affected, suggesting a systemic shift.

Key Perspectives

Open source maintainers (Madhavapeddy, Craig-Wood): Facing an overwhelming volume of disclosures — Craig-Wood's project saw a 40-fold monthly increase. Maintenance burden is spiralling, even with AI triage tools. Embargo practices are broken.

Security researchers and AI developers: The ability to find exploits from minimal hints is a demonstration of powerful AI capabilities, but it also raises ethical questions about the design of coding agents. Some models (like Claude Fable) can refuse harmful tasks; others (like DeepSeek V4 Pro) may not.

Critics/Skeptics: Some may argue that the reports are anecdotal and that not all projects see the same surge. Others might point out that automated exploit discovery also helps defenders by finding bugs faster — if the disclosure process is adapted. The risk is that the response is reactive rather than systemic.

What to Watch

The number of security disclosures reported by prominent open source projects over the coming months, to see if the surge is sustained or an anomaly.
GitHub's CVE assignment turnaround time — if it continues to stretch to weeks, it signals capacity strain.
Whether foundations like the Open Source Security Foundation (OpenSSF) or the Linux Foundation propose new embargo frameworks or automated disclosure tools.

ZOTPAPER