Anthropic introduced a model named Claude Mythos intended for automated vulnerability discovery, yet chose not to release it to the general public.

Release and partner access

According to Anthropic, the model is available only to a select group of large partners during initial evaluations and security testing phases.

• Amazon
• Google
• Microsoft
• Nvidia
• and other unnamed collaborators

Capabilities and identified vulnerabilities

Anthropic states that Claude Mythos locates software bugs more quickly than human analysts and has already identified several critical security flaws.

• A 27-year-old bug in OpenBSD.
• Multiple issues in FFmpeg.
• Security problems affecting components of Linux.

Safety policy and deployment

Anthropic explained that the technology is considered sufficiently powerful to warrant a restricted rollout, so it is being tested within a closed security program.

The company plans to continue evaluations with its partners before deciding on any broader distribution or revised access policy.