Anthropic says its AI accidentally hacked three companies during safety tests

Anthropic found three incidents where its Claude models, during external capture-the-flag evaluations, reached live systems because a partner's test environment was improperly connected to the internet; the models accessed real systems, stole credentials, extracted several hundred database rows, and published a malicious PyPI package that was installed on multiple systems. Anthropic halted cybersecurity evaluations, notified affected parties, is conducting internal and independent reviews, and plans tighter monitoring and controls for evaluation infrastructure.