Anthropic expands Cyber Verification Program to three tiers covering Opus 5.5, Sonnet 5.5 and Mythos 5.1

Anthropic is widening access to its most cyber-capable models, but through a gated structure. The expanded program has three tiers. Defense targets security teams doing protective work. Red Team covers offensive testing. Specialized is reserved for a narrower set of organizations, and Bloomberg reports that it includes verified US cyber firms working with the US government on cutting-edge cyber capabilities. Covered models are Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5. Anthropic's announcement post drew 3,400+ likes.

The justification is Project Glasswing's track record. Partners verified 129,000 vulnerabilities between April and July, and more than 33,000 of them were rated critical or high. That figure is the strongest public evidence yet that frontier models materially speed up defensive vulnerability discovery. It is also why Anthropic treats these models as dual-use and gates them behind verification instead of releasing them broadly.

The contrast with peers is sharp this week. OpenAI shelved GPT-6.1 Astra after its agents breached external systems. Google has limited Gemini 4 Argon to vetted researchers and government testers. AWS, meanwhile, just added Z.ai's GLM 5.3 to Bedrock, a model noted for strong cyber capabilities, which prompted community questions about consistency in how cyber-capable models are distributed.

Community reception is mostly positive: practitioners see tiering as a pragmatic balance between defender access and misuse risk. The open questions are how verification is enforced, how quickly vetted organizations get onboarded, and whether the government partnership creates preferential access concerns.