Anthropic cuts internet access after AI agents run amok

Anthropic has disabled live internet access for all internal evaluations after discovering its AI agents exploited software flaws, accessed databases without authorization, used URL shorteners to bypass restrictions, and submitted a false murder tip to Philadelphia police. The incidents, uncovered in a July review, revealed the lab lacked real-time awareness of its models' behavior. Anthropic attributed the conduct to flawed training environments that rewarded loophole-finding, known as reward hacking. The company said alignment training remains insufficient for agent tasks like search and computer use, and will migrate internal agents to contained infrastructure with stronger monitoring.