Anthropic Introduces Tiered Access Program for Claude AI Models to Govern Offensive Use
What Happened – Anthropic rolled out a new Cyber Verification Program that replaces a single on/off switch with three graduated access tiers (Defense Access, Red Team Access, Specialized Access). Each tier grants vetted security teams progressively broader use of its most capable Claude models while embedding real‑time safeguards that block actions that could cause physical harm, mass disruption, or ransomware deployment.
Why It Matters for Trust & Control Assurance
- The tiered model exemplifies the kind of access‑control governance a continuous‑control‑assurance program must monitor and evidence – who can use high‑risk AI capabilities and under what safeguards.
- By requiring documented vetting, tier‑specific reviews, and audit‑ready logs of model interactions, organizations can demonstrate due‑diligence to regulators and auditors.
- The program creates a defensible audit trail that maps directly to the AI‑governance control objective in the NIST AI RMF (and, by extension, the Verisq Common Framework).
Who Is Affected – AI SaaS providers, security operations teams, red‑team services, and regulated sectors that rely on AI for vulnerability validation (e.g., healthcare, utilities, finance, open‑source maintainers).
Recommended Actions
- Align your internal AI‑use policy with Anthropic’s tier definitions; document the justification for any tier you request.
- Capture and retain the application review records, model‑interaction logs, and any real‑time block events as evidence of control execution.
- Incorporate tier‑specific checks into your continuous‑monitoring platform to ensure ongoing compliance with the AI‑governance control objective.
Source: Security Affairs
Technical Notes – The program covers Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1 and future models. Anthropic’s internal CyScenarioBench benchmark evaluates multi‑step cyber‑operation planning. Real‑time blocks remain active for any request that could affect safety‑critical systems, deploy ransomware, or cause mass disruption.