Anthropic withholds Claude Mythos, launches Glasswing

This is both a safety move and a strategic moat move: Anthropic is framing capability restraint as responsible deployment, while turning its strongest model into a controlled enterprise/security product.

• –If Mythos really finds zero-days at this scale, public release would be a liability as much as a feature
• –Project Glasswing gives Anthropic a way to monetize frontier cyber capability without handing attackers the same tooling
• –The partner list matters: access through AWS, Vertex AI, and Microsoft Foundry makes this feel like platform distribution, not just a research demo
• –The competitive angle is obvious too: if Anthropic can own “defensive-only frontier model” positioning, that becomes a differentiated story versus general-purpose releases
• –The real test is whether the model produces durable defensive outcomes, not just flashy vulnerability counts