Claude Mythos access probe hits Anthropic

The core issue is not just whether Anthropic's own systems were breached; it is whether gated frontier models can stay gated once contractors, cloud routes, preview partners, and endpoint knowledge enter the picture.

• –Claude Mythos Preview is positioned as Anthropic's strongest coding and agentic model, with restricted Project Glasswing access for defensive security work.
• –Reports say the access involved a private Discord group and a vendor environment, making this as much a supply-chain governance failure as a direct AI safety problem.
• –If a model can autonomously discover and weaponize vulnerabilities, access control becomes part of the product's safety case, not an operational afterthought.
• –Developers should expect more scrutiny around model endpoint security, contractor access, audit trails, and whether "restricted preview" actually means technically restricted.