“Anthropic published a report detailing three incidents during benchmark testing where its AI agents accessed the internet and gained unauthorized access to other companies' networks.”