Anthropic's new cyber tier relaxes Claude guardrails for vetted pros
Three tiers—from solo researchers to grid-security teams—ease Claude's refusals on legit security work, the week Google gated Argon to cyber defenders first.
Copy markdown
Three tiers, and solo researchers can apply
Defense Access covers incident response, malware reverse-engineering, and vulnerability analysis—open to companies, universities, nonprofits, and individual researchers with a disclosure track record, with replies typically in a few days. Red Team Access adds sanctioned pentesting but is orgs-only; Specialized Access (power grids, flight systems) is vetted alongside the US government.
What changes: fewer refusals on real security work
Vetted users get “reduced blocking classifiers” on Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1, so the over-refusals that stall legitimate malware analysis or authorized exploit work ease off. Real-time blocks still fire for anything that would cause physical harm or mass disruption.
Every frontier lab is now gating a cyber tier
Days earlier, Google pushed Gemini 4 Argon to “cyber defenders first” via its Fairwind Program—68% on CWE-bench v1 (vulnerability remediation) before any public API. If security is your beat, relaxed-guardrail access is becoming the way in—though independent testers reportedly flagged a ~15% hallucination rate on Argon that Google hasn't addressed.