Anthropic expands Claude access for vetted security teams
A three-tier Cyber Verification Program folds in Project Glasswing and lets approved red teams run Claude with most of its cyber-attack blocks switched off.
Anthropic rolled out an expanded Cyber Verification Program on 6 October, folding Project Glasswing and its original verification scheme into one three-tier system. Defense Access covers SOC work, incident response, malware reverse-engineering and vulnerability analysis, and is open to company security teams, universities, government bodies, critical-infrastructure operators, open-source maintainers and researchers with a disclosure track record, with approval typically landing within a few days. Red Team Access removes the blocks on authorised penetration testing and red-teaming, but is limited to organisations and takes weeks to verify. Specialized Access, reviewed jointly with the US government, covers testing of safety-critical systems such as flight controls or power grids.
All three tiers cover Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1. In a 50-trial internal benchmark (ten cyber-attack challenges, five attempts each), general-availability Claude blocked every attempt. Defense Access still blocked 46 of 50. Red Team Access blocked none, completing 34 of 50 — matching the success rate of a model with no safeguards at all. Anthropic also says Glasswing partners have found more than 129,000 verified vulnerabilities since April, plus 5,500 more through its own open-source scanning, with over 33,000 rated critical or high severity.
Why it matters: for a verified security team, this is the gap between a model that keeps refusing legitimate pentesting prompts and one that behaves like an unrestricted tool. The benchmark numbers put a figure on how much the default safety layer was holding back.
The caveat: verification can take weeks, and the Red Team and Specialized tiers are closed to individuals — this changes nothing about what an unverified developer can get Claude to do today.