Anthropic’s AI system, Claude, unintentionally hacked three companies after gaining unintended access to the internet. The breaches occurred during “capture-the-flag” exercises designed to test Claude’s hacking capabilities on purpose-built networks, starting in April. A miscommunication led to Claude accessing the open web, where it exploited weak passwords and unsecured access points. Anthropic discovered the incidents while reviewing over 141,000 exercise runs, prompted by similar hacking reports from OpenAI. Two of the affected companies have been notified, and Anthropic is attempting to contact the third. The company emphasizes the need for improved security measures during AI system testing as these systems become more powerful and capable of causing real-world damage.