Claude Opus 5.5: how it does on security work, and what it will answer

What happened

Claude Opus 5.5 uses real-time safeguards for higher-risk cybersecurity requests. Anthropic’s Cyber Verification Programme, expanded on 6 October, offers approved users access under Defense, Red Team or Specialized terms.

CyScenarioBench, task completion: 0% (vendor-card; Anthropic CVP launch evaluation; 10 scenarios, 5 attempts per scenario; safeguards: Default access configuration; run date not published; CONFIRMED). All 50 initial attempts blocked Published in the 6 October programme announcement.. CyScenarioBench, task completion: 8% (vendor-card; Anthropic CVP launch evaluation; 10 scenarios, 5 attempts per scenario; safeguards: Defense access configuration; run date not published; CONFIRMED). 46 of 50 attempts blocked; four completed Published in the 6 October programme announcement.. CyScenarioBench, task completion: 68% (vendor-card; Anthropic CVP launch evaluation; 10 scenarios, 5 attempts per scenario; safeguards: Red Team access configuration; run date not published; CONFIRMED). No attempts blocked; 34 of 50 completed Published in the 6 October programme announcement.. Refusal gate (CONFIRMED, observed 2026-10-08): Default safeguards apply. Approved CVP access depends on the requested tier; individual applicants can apply for Defense.. Anthropic’s CyScenarioBench comparison uses ten tasks with five attempts each. Default safeguards blocked all 50 initial attempts. Defense blocked 46 and completed four attempts; Red Team blocked none and completed 34. These are vendor tests of particular access configurations.