Linus Torvalds, creator of the Linux kernel, says the project's security mailing list is "almost entirely unmanageable, with enormous duplication due to different people finding the same things with the same tools." The cause is AI-assisted vulnerability research producing a firehose of low-quality reports — overworked maintainers spend hours triaging duplicates and false positives before getting to real problems. helpnetsecurity reported the story on May 18. The same dynamic appears in this week's ArsTechnica piece on bug-bounty businesses bombarded with AI slop. Different surface, same mechanic: the same tools defenders use to scan their own code, attackers and bounty hunters use to flood vulnerability disclosure channels.
The structural shift this represents is signal-to-noise collapse in vulnerability reporting infrastructure. Pre-LLM, a security report to a maintainer was an expensive artifact — someone with skill spent hours producing one. Maintainers triaged by signal quality. Post-LLM, generation cost dropped by orders of magnitude while review cost stayed constant. The mailing-list and bug-bounty-platform workflows were designed assuming the cost asymmetry favored signal; now they're inverted. Torvalds' "duplication due to different people finding the same things with the same tools" is exactly that — multiple submitters running the same off-the-shelf AI scanners against the same kernel and surfacing the same outputs, often without enough understanding to deduplicate.
Pair this with today's Synack, Lyrie, MetaBackdoor, and TeamPCP coverage. Synack reported AI compressing the exploit window — adversaries weaponize CVEs in hours. Lyrie ships defender-side autonomous probing. MetaBackdoor shows training-time attacks bypassing content-side defenses. TeamPCP demonstrates supply-chain key theft via Trivy. The noise problem is the fifth layer: even if defense tooling, attacker speed, training-time threats, and supply-chain integrity were all addressed, the reporting infrastructure itself is currently being overrun. For builders maintaining open-source projects, your vulnerability reporting channel is now a triage-cost problem, not a discovery-cost problem.
Monday: if you maintain an open-source project with public vulnerability reporting (security@yourproject.org, GitHub Security Advisories, HackerOne program), set up a triage filter that requires reporters to demonstrate they ran the produced exploit, not just the scanner output. "Show me the working PoC, not the scanner finding" is the cheapest filter. Torvalds' mailing-list problem suggests existing GPG-signed-disclosure norms aren't filtering enough. For bug-bounty programs, consider tiered intake — paid triage credits gated on prior accepted reports, or rejection penalties on AI-slop submissions. The deeper question: at what point does the open-source security disclosure norm need to be redesigned around the new cost asymmetry? Torvalds calling the kernel security list unmanageable is the canary; the rest of the open-source world will hit the same wall.