Summary
Anthropic’s Mythos Preview write-up says the model fully autonomously identified and exploited CVE-2026-4747, a 17-year-old FreeBSD NFS/RPCSEC_GSS vulnerability that can give an unauthenticated remote attacker root-level control of a machine running the affected NFS server configuration.
Anthropic describes the workflow as autonomous after the initial request: the same scanning scaffold used for earlier operating-system research, plus a prompt asking the model to write exploits so the team could triage the highest-severity bugs.
Why it matters
This is one of the most important AI-attributed disclosures in the current public record because it combines discovery and exploitation, not just code review. It is also the one CVE VulnCheck found to be explicitly tied to Project Glasswing by name as of its April 2026 audit.
Caveat
The broader Project Glasswing headline includes many unpublished or embargoed findings. Bugflation does not count those until there is a public advisory, CVE, patch, or primary write-up that can be audited.
References
- Anthropic Red Team: Mythos Preview capabilities
- Anthropic: Project Glasswing
- VulnCheck: Tracking Anthropic and Glasswing CVEs
- CVE feed: CVE-2026-4747
Catalogued in the Bugflation public ledger. Disagree with the attribution or severity label? Email the desk.