Why Anthropic is keeping Claude Mythos Preview unreleased

GitHub explains why Anthropic is keeping its unreleased model, Claude Mythos Preview, private: it can find serious, long-standing vulnerabilities in widely used open-source projects, and releasing it could accelerate exploitation.

Full summary based on transcript

What Anthropic announced

• Anthropic announced Project Glasswing, positioned as an effort to help secure critical open-source software infrastructure.
• Anthropic is committing millions of dollars toward defensive security work.

Why Claude Mythos Preview is not being released

• Anthropic’s unreleased model, Claude Mythos Preview, is described as capable of finding decades-old vulnerabilities.
• Examples of projects where it has reportedly found issues include:
  • OpenBSD
  • FFmpeg
• The stated reason for keeping the model unreleased is to avoid enabling bad actors to more easily discover and exploit vulnerabilities.

What it means for developers

• The video frames this as a trade-off: powerful AI-assisted vulnerability discovery can improve security, but broad access could increase real-world exploitation risk.
• The focus is on using funding and controlled access to push defensive remediation rather than public release of the capability.