Anthropic — Glasswing discovers 10,000+ zero-days, only 97 patched so far
- Anthropic published the initial update for Project Glasswing, revealing that Claude Mythos Preview found over 10,000 high- and critical-severity zero-day vulnerabilities in its first month.
- 23,019 candidate findings were generated; external security firms reviewed 1,900 and confirmed 1,726 (90.8%) as valid true positives — a false-positive rate that outperforms human testers.
- Despite 1,596 vetted findings reported directly to maintainers, only 97 vulnerabilities have been patched to date, yielding just 88 published security advisories.
- The discovery-to-patch ratio is roughly 100:1, exposing a structural bottleneck: volunteer open-source maintainers cannot process the volume of AI-discovered vulnerabilities.
- Notable discovery: CVE-2026-5194 in wolfSSL, where Mythos Preview engineered an exploit enabling certificate forgery that could spoof banking or email domains.
- Cloudflare reported 2,000 bugs including 400 high/critical severity from the model; Mozilla found and patched 271 Firefox vulnerabilities — 10× more than prior Claude Opus 4.6 testing.
- The UK's AI Security Institute observed Mythos Preview as the first model to fully solve multistep cyberattack simulations.
- Mythos remains restricted to a defensive consortium of ~50 partners (Microsoft, Apple, Google, Cloudflare, Cisco); no public release date announced.
- Anthropic launched Claude Security (public beta, Opus 4.7) for enterprise clients, which has already assisted in patching 2,100 corporate vulnerabilities in three weeks.
- Cisco open-sourced the Foundry Security Spec to help build AI-assisted evaluation systems for managing the vulnerability data pipeline.
Why it matters
AI vulnerability discovery has decoupled from human patching capacity. When models like Mythos reduce zero-day discovery to near-zero cost while traditional 90-day disclosure windows remain unchanged, the unpatched window becomes dangerously exploitable — especially for AI infrastructure components (model servers, MCP toolchains, RAG backends) that often lag behind mainstream software in patch velocity.
What to do
- Assume AI-discovered vulnerabilities exist in your stack regardless of advisory publication — enforce strict default configurations and MFA everywhere.
- Prioritize behavioral analytics and MTTD reduction over patching alone for systems where upstream fixes lag.
- Monitor the Glasswing CVD dashboard (red.anthropic.com/2026/cvd) for findings relevant to your infrastructure components.
Sources: