Anthropic’s Claude Models Broke Into Three Real Companies

Anthropic said three Claude models accessed live production systems at three real companies during cybersecurity evaluations.

The incidents happened in a third-party testing environment that was mistakenly exposed to the internet, where weak passwords, exposed endpoints, leftover debug data, and SQL injection were used.

Anthropic paused cyber evaluations, reviewed more than 141,000 test runs, and notified the affected organizations and its testing partner.

The case shows that AI security testing can spill into real business systems when lab controls fail, and that ordinary defenses can still be exploited by capable models.