AI found 23,000 software bugs. The humans can't keep up.

Let's do the numbers, properly. 1,000+ open-source projects scanned. 23,019 total findings. Of those, 6,202 rated high or critical severity. Here's the bit the breathless coverage flattens: the 90.6% validation rate isn't a stamp on all 6,202. Six external firms assessed a sample of 1,752 high/critical findings, of which 90.6% (1,587) were valid true positives and 62.4% (1,094) confirmed high or critical. That's still a strong number — but it's a sample, and the difference matters if you're going to quote it. Restricted to roughly 50 partners — AWS, Apple, Google, Microsoft, NVIDIA, JPMorgan, Cloudflare, Mozilla — whose own bug-finding reportedly jumped more than tenfold (Cloudflare alone reported ~2,000 bugs). As capability demos go, this one is hard to wave away. The headline is earned; just don't round the asterisk off it.

Anthropic basically admitted it in the same update: finding bugs is now cheap; fixing them is not. A lot of critical open-source software runs on the goodwill of a handful of unpaid volunteers who already can't keep up with normal patch demands. Point an automated vulnerability-finder at those codebases and you get thousands of confirmed, serious flaws — and a patching queue no volunteer army can clear at pace. The wolfSSL certificate-forgery flaw (CVE-2026-5194, CVSS 9.1) was a clean example: Mythos didn't just spot it, it built a working exploit to forge certificates; responsible disclosure followed and the maintainers shipped a fix. But wolfSSL has dedicated resources. The 27-year-old OpenBSD flaw and the 16-year-old FFmpeg one Mythos also surfaced tell you how long this stuff hides in plain sight — and most critical open-source projects don't have wolfSSL's funding. The bottleneck Anthropic flagged isn't a temporary operational hiccup. It's structural — Anthropic's own dashboard shows high/critical bugs averaging ~two weeks to patch and maintainers asking it to slow disclosure.

Now imagine the same tool, not restricted, in offensive hands. You don't need to speculate. Mythos flagged 6,202 high/critical vulnerabilities that were previously unknown — and in at least one case (wolfSSL) it built a working exploit, not just a flag. The same capability that helps a defender write a patch helps an attacker write an exploit — and exploits are faster to write than patches are to deploy, test, review and ship to production. The asymmetry in discovery is real and documented. The same asymmetry, flipped, is the attack-surface reading of this paper.

This is the detail the press releases don't headline. Anthropic's restriction of Mythos Preview to roughly 50 vetted partners is a policy decision, not a technical one. The model itself is not architecturally constrained from running offensive analysis. The access gate is an agreement with a vetted list — a list that presumably involves contracts, NDAs, and trust assessments. That is meaningful; institutional controls are not nothing. But it's a different kind of guarantee than 'the model literally cannot help you write an exploit'. At some point, either the capability becomes widely available (because other labs build similar systems, or because this one leaks), or the policy holds. Which of those you think is more likely probably determines how worried you are right now.

So file this under 'capability shift, confirmed' — and under 'the bottleneck moved'. For a decade the hard part of security was finding the holes. Mythos says that era is ending. The new hard part is the boring, human, under-funded work of closing them before someone else walks through. Anthropic's OpenSSF partnership, Claude Security beta and free tooling are an acknowledgement of that — a welcome one. But notice it didn't put a number on the money, and the patch dashboard is already running behind. Whether the support is large enough, or fast enough, to keep pace with the discovery engine they just demonstrated is a different question — and one the update does not answer.