A researcher documents AI-driven inversion of exploit timing

Cambridge computer scientist Anil Madhavapeddy wrote that automated exploitation was, by mid-2026, arriving faster than the process meant to precede it: after seeing exploit probes appear within roughly ten minutes of a public discussion about a vulnerability, he built a working exploit himself in under a minute, and argued “mean time to exploit” — the gap between disclosure and attack — had gone negative, with attacks sometimes arriving before or alongside a patch.

He cited two examples: CVE-2026-39987, in the notebook tool Marimo, saw its first exploitation attempt roughly nine hours after the advisory; CVE-2026-33017, in the low-code tool Langflow, was exploited within about twenty hours. He pointed to corroborating experience from Nick Craig-Wood, maintainer of the file-sync tool rclone, who reported over 40 vulnerability disclosures in a single month against roughly 20 across the prior decade — a shift Madhavapeddy attributed to coding agents now routinely scanning open-source projects for flaws — and cited a study by Fang et al. finding coding agents exploited 87% of vulnerabilities when given a CVE’s description, against 7% with none.

The argument was about incentives rather than any single flaw: embargoed disclosure assumes defenders have a head start measured in days, and Madhavapeddy argued that no longer reliably holds once one public mention can point an automated agent at the problem. It added a maintainer’s-eye view to a summer of lab-side warnings about agentic cyber-offence, including OpenAI’s disclosed autonomous-agent breach of Hugging Face and Greg Brockman’s “Defender’s Window” essay.