Anthropic to widen Claude Mythos access amid cyberattack concerns

Anthropic said it will expand customer access to its Claude Mythos models “in the coming weeks” after completing additional tests, announcing the timeline alongside the release of a new Opus 4.8 model. The company added that it is developing safeguards and expects broader availability soon.

Access to Mythos has been limited to a controlled program called Project Glasswing, which granted use to select technology companies, security researchers and government partners while Anthropic conducted a restricted preview.

Security researchers and government agencies have raised concerns about Mythos because tests showed it can identify software vulnerabilities and complete multi-step simulated attacks without human guidance. A U.K. security institute reported that the model finished a 32-step simulated corporate network attack in testing. In separate internal evaluations, engineers found the model flagged hundreds of potential flaws in major software, including a review that identified 271 issues in a popular browser. A security startup reported that a preview version of the model helped researchers produce an exploit chain against Apple’s M5 Mac chips.

Anthropic has said Mythos can speed defensive work by helping find and patch vulnerabilities before attackers do. The company previously described Mythos as “the most powerful AI model we’ve ever developed” and positioned it above its Opus series.

The planned wider rollout has prompted debate across the AI industry over how advanced systems should be released and regulated. OpenAI CEO Sam Altman criticized Anthropic on a podcast, describing the firm’s cybersecurity warnings as “fear-based marketing.”

Anthropic did not specify which safeguards remain incomplete, how testing will be expanded, or whether broader availability will include the same level of functionality for all customers. The company did not immediately respond to requests for further comment.