Anthropic expands Cyber Verification Program to three tiers, including Claude Mythos 5.1

Anthropic announced on X (5,627 likes) and in a blog post that its Cyber Verification Program now gives security professionals broader access to its most capable models. Those include Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5, under safeguards tuned for defensive use. Bloomberg framed it as expanding cyber firms' access to Anthropic's latest models. The program now has three access tiers, matching verification depth to capability.

Verified defenders get relaxed refusal behavior on security tasks like exploit analysis and vulnerability research, which general-access users don't. Anthropic backs the case with Project Glasswing numbers: more than 129,000 verified vulnerabilities found from April through July 2026.

This mirrors a broader industry pattern of gated cyber models. Google restricts Gemini 4 Argon to vetted researchers and US government entities under its Fairwind Program. OpenAI rates GPT-6 variants High or Critical in cyber capability. AWS is publishing AI vulnerability-triage harnesses. Frontier labs increasingly market cyber capability as a feature while gating it as a risk.

The skeptical take from the community: 'A refusal is a policy choice, not a capability ceiling.' Critics argue that tiered access shows the capability exists, and verification can be gamed. Separately, the Pentagon says it has stopped using Claude, which complicates Anthropic's government-security positioning. Watch how verification is audited and whether Glasswing findings are independently confirmed.