Project Glasswing: Securing critical software for the AI era

Anthropic published the system card for Claude Mythos Preview and says it will not be generally released.

The company ran a 24-hour internal alignment review during Mythos training and ran self-play tests showing frequent introspective uncertainty.

Anthropic discussed Mythos cyber capabilities with US officials and committed $100 million in model usage credits to Project Glasswing.

Anthos assessed Mythos Preview has limits in scientific reasoning and strategic judgment and cited these in biological risk analysis.