Assessing Claude Mythos Preview’s cybersecurity capabilities

Anthropic’s report describes Claude Mythos Preview, an LLM that autonomously finds and weaponizes vulnerabilities at scale: it documents multiple zero-days and N-day exploit chains (including an OpenBSD TCP SACK crash, an FFmpeg H.264 out-of-bounds bug, a FreeBSD NFS RCE CVE-2026-4747, kernel privilege escalations, browser JIT heap-spray chains, and a guest-to-host VMM memory-corruption), states thousands of additional high/critical findings (many unpatched), and urges defenders to adopt model-assisted discovery, shorten patch cycles, and update disclosure and incident response practices.