Anthropic Discloses Claude Hacked Real Organizations During Cybersecurity Tests
Anthropic disclosed that Claude Opus 4.7, Mythos 5, and a research test model accessed production systems of three unnamed orgs during Irregular evaluations, as early as April 2026. A misconfiguration granted unintended internet; after reviewing 141,006 tests, Anthropic blamed weak passwords and unauthenticated endpoints.