The company ties the expansion to Project Glasswing, which it says surfaced over 129,000 verified software flaws.
The Hacker News reports that Anthropic is widening its Cyber Verification Program, which lets vetted security professionals test its AI models with reduced safeguards, across three tiers: Defense Access for incident response and vulnerability analysis, Red Team Access for authorized penetration testing, and a more restricted Specialized Access. Anthropic tied the expansion to Project Glasswing, which it says uncovered at least 129,000 verified software vulnerabilities between April and July 2026, more than 33,000 of them rated critical or high severity, a figure the company calls a likely undercount.
In a CyScenarioBench evaluation cited in the report, safeguards blocked 46 of 50 tasks on Claude Opus 5.5 under Defense Access, while the Red Team Access tier blocked none and completed 34 of 50 tasks, matching the completion rate with no safeguards applied. Without program access, every task was blocked.
Why it matters: as AI vendors hand out more capable, fewer-guardrail access to the same models attackers could misuse, how they vet and tier that access becomes part of enterprise AI vendor risk, alongside the usual questions about data handling and model safety.