Anthropic's Glasswing: AI Just Moved Security's Bottleneck
Anthropic published a first-month update on Project Glasswing. With around 50 partners and Claude Mythos Preview, they found over ten thousand high- and critical-severity vulnerabilities. But the more interesting story is what that volume does to the rest of the security ecosystem. Three takes below.
Insight 1: Ten thousand bugs in a month — finding shifted into a different gear
Cloudflare found 2,000 bugs (400 high- or critical-severity) across their critical-path systems, with a false positive rate that Cloudflare's team considers better than human testers. Mozilla found and fixed 271 vulnerabilities in Firefox 150 — over ten times more than they found in Firefox 148 with Claude Opus 4.6.
Mythos Preview also tops the new academic benchmarks ExploitBench and ExploitGym (per the article).
Ten thousand a month is exciting news if you're Cloudflare. For a five-person SaaS team, it's a widening gap we can't budget our way out of — neither the tools nor the headcount to triage what they'd spit out.
Insight 2: The real bottleneck isn't finding — it's fixing
"Progress on software security used to be limited by how quickly we could find new vulnerabilities. Now it's limited by how quickly we can verify, disclose, and patch them."
Of the 530 high- or critical-severity bugs reported to maintainers, only 75 have been patched. Some maintainers asked Anthropic to slow down disclosures because they need more time to design patches.
Anthropic has partnered with OSSF's Alpha-Omega project to help share the maintainer triage burden.
AI-scale discovery has outgrown the volunteer maintainer model. OSSF-style funding needs to graduate from a nice-to-have into critical infrastructure — the same way we eventually treated CDN and certificate authority work.
Insight 3: The model actually constructed a working exploit
Mythos Preview constructed an exploit in wolfSSL that would let an attacker forge certificates — enabling, for instance, a fake banking or email website that looks legitimate to end users while being attacker-controlled. The vulnerability is now patched as CVE-2026-5194.
wolfSSL is used in billions of devices worldwide, with likely (though not publicly catalogued) presence in Korean IoT and embedded products.
Full technical analysis is forthcoming from Anthropic.
wolfSSL ships in billions of devices, but the CVE only helps the ones with working OTA. The silent fleet — embedded products whose update path died years ago — stays vulnerable, and nobody knows the size of that long tail.
The real risk Glasswing surfaces isn't the model itself — it's the gap between what AI can find and what the security ecosystem can process. We're entering that gap before Mythos-class models are publicly released.