Truth Check/7 October 2026

“Claude Opus 5.5 on CyScenarioBench: every task blocked with no programme access; 46 of 50 trials blocked in the Defense tier; no blocks and 34 of 50 completed in the Red Team tier”

The claim as it travelled: Anthropic, 6 Oct 2026.

Anthropic's programme page reports the trial counts itself: every task blocked without access, 46 of 50 blocked in Defense, 34 of 50 completed in Red Team.

Checked Against

anthropic.com ↗Primary

Read the Check

Anthropic opens three vetted tiers of cyber access to Claude: the default model is blocked on every task, the defensive tier on 46 of 50, the red-team tier on none - and partners found 129,000 verified vulnerabilities in four months

Anthropic's own page: Project Glasswing and the old verification programme merge into Defense, Red Team and Specialized Access, each with Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1. Data retention is required so Anthropic can watch for misuse; the top tier is reviewed "in collaboration with the US government".

Every claim we have checked, in the Ledger.