Anthropic restricts Claude Mythos over zero-day exploits

Anthropic withholding a frontier model due to offensive capabilities marks a turning point where AI safety concerns are actively dictating product releases.

• –Mythos operates at a new tier above Opus, finding bugs that survived millions of automated tests, including a 27-year-old OpenBSD vulnerability.
• –"Project Glasswing" essentially turns the model into a privatized cyber-defense asset for partners like AWS, Microsoft, and CrowdStrike.
• –The restricted release validates fears of autonomous cyber-agents, proving frontier models can function as zero-day discovery engines.
• –This decision will ignite massive debate over whether private AI labs should act as the gatekeepers of weapons-grade tools.