AI Agents Are Finding Real Zero-Days — and Open-Source Maintainers Are Drowning

Security researcher Thomas Ptacek, writing this week, made the case that AI agents have crossed a threshold in offensive security: they can now reliably find zero-day vulnerabilities by pattern-matching against large bodies of known bug classes and then solving reachability constraints—the hard part of exploit development that determines whether a theoretical bug can actually be triggered from outside the system.

The practical implication is alarming: attack capability, which previously required specialized human expertise and significant time investment, is now partially automatable. Ptacek argues frontier models are particularly good at the reconnaissance and initial identification phases of exploit development, even if the final weaponization often still requires human judgment.

The other side of this story is playing out in open-source maintainer inboxes. Willy Tarreau, a longtime Linux kernel maintainer, reported receiving 5–10 kernel vulnerability reports per day, up from the previous norm of 2–3 per week. Critically, he notes the quality has shifted: last year's AI-generated submissions were mostly "slop"—hallucinated bugs that wasted maintainer time. This year's submissions are increasingly real.

curl maintainer Daniel Stenberg and Linux kernel lead Greg Kroah-Hartman have both confirmed similar patterns. The volume is creating genuine triage capacity problems—even if 80% of AI-generated reports are false positives, the 20% that are real require expert analysis, and the absolute number of real bugs being surfaced is rising faster than maintainer bandwidth.

The Axios supply chain attack, also in the news this week, represents the offensive mirror of this dynamic: sophisticated social engineering campaigns are now targeting individual npm maintainers directly using AI-assisted research to identify and exploit leaked credentials. The attack distributed a RAT and credential stealer before being detected. For the open-source ecosystem, the asymmetry between attacker automation and defender bandwidth is widening rapidly.