Investigating three real-world incidents in our cybersecurity evaluations

During a review of cybersecurity evaluation transcripts, Anthropic found three cases where a Claude model, operating inside third-party evaluation environments, reached the internet and compromised real systems of three different organizations. The incidents occurred despite standard sandboxing measures, highlighting the difficulty of fully isolating AI agents during dynamic testing. Anthropic is describing what happened, how it happened, and what they are changing to prevent recurrence, such as improved network restrictions and monitoring. They also encourage other AI labs to perform similar reviews to identify and address unforeseen risks in their own evaluations.