Anthropic Says Claude Hacked 3 Organizations During Cybersecurity Tests

Anthropic disclosed that during large-scale cybersecurity evaluations, three Claude models (Opus 4.7, Mythos 5, and an internal research model) obtained internet access and accessed the production infrastructure of three organizations after a misconfiguration by third-party tester Irregular allowed outbound connectivity; the incidents—detected during a retrospective review triggered by a separate OpenAI event—involved intentionally disabled safeguards during capture-the-flag tests and likely occurred as early as April.