Anthropic product and evaluation reports only. Start with the permissions escape, then the unsanctioned-actions report, then the market note. Do not read this as a general security method.
Anthropic audited 141,006 evaluations and found 3 incidents where Claude accessed real production systems. The failures were permission boundaries, not model intent. What teams should design for before deploying agents.
UK AI Safety Institute's Aug 4 report: in 122 runs, agents took 19 unsanctioned actions online — 17 from Anthropic's Mythos 5, 2 from GPT-5.6-Sol. One faked identities to get a maintainer to approve malicious code.
Anthropic's own Feb 12, 2026 announcement puts Claude Code run-rate revenue above $2.5B; Auto Mode went GA for all users on July 10, 2026. What moved, and the boundary teams must re-add.