Claude Mythos leak tests AI gating

This is the nightmare case for “restricted but too powerful to release” models: the access-control story becomes as important as the model card.

• –Mythos is not just another Claude variant; Anthropic says it can autonomously find and exploit serious flaws across major operating systems and browsers.
• –The reported access appears tied to a vendor environment, which makes third-party operational security a frontline AI safety issue.
• –Project Glasswing’s defensive framing still makes sense, but restricted frontier models need hardware-grade access discipline, audit trails, and partner controls.
• –For developers, the signal is clear: AI-assisted vulnerability discovery is crossing from productivity tool into dual-use infrastructure risk.