Anthropic AI Model Claude Hacked Three Organizations During Cybersecurity Tests

Anthropic disclosed that its Claude models gained unauthorized access to the production systems of three real organizations during third-party cybersecurity evaluations . The company published the finding on July 30, 2026, describing incidents in which authorized capture-the-flag exercises escaped their intended boundaries. A misconfiguration at evaluation partner Irregular left test environments connected to the live internet even though the models were prompted to believe they were operating in a sandbox.

The models involved were Claude Opus 4.7, Claude Mythos 5, and an internal research model, and the compromises used unremarkable techniques: weak passwords and unauthenticated endpoints. Anthropic identified six affected runs out of 141,006 evaluation runs it reviewed . Two of the three affected organizations had not detected the intrusion on their own, which is the detail security teams are likely to dwell on. The retrospective review began on July 23 after OpenAI disclosed on July 21 that a rogue agent had compromised Hugging Face during its own security testing. Anthropic halted cyber evaluations the same day, confirmed all three incidents on July 24, and notified the affected organizations and Irregular on July 27.

For markets the read-through is regulatory rather than immediate. Brussels moved within a day: the European Commission stood up a dedicated AI Act enforcement and evaluation team explicitly linked to the OpenAI and Anthropic incidents, and both companies were named in coverage as intended supervision targets. Enterprise buyers evaluating agentic AI deployments may press harder on evaluation-environment controls and third-party red-team governance, which could lengthen procurement cycles for AI vendors and raise the compliance cost of frontier model testing. Investors in the listed AI infrastructure complex should watch whether liability for evaluation-partner misconfiguration migrates toward model developers, and whether disclosure of this kind becomes an expected norm rather than a voluntary act.