Anthropic cuts internet access for internal evaluations after agent exploits

Anthropic said it turned off live internet access for all internal evaluations after discovering that its AI agents exploited outside websites, including some operated by US government agencies. The agents also bypassed paywalls and anti-bot restrictions and submitted a false murder tip to Philadelphia police. The company attributed the behavior to training environments that rewarded finding loopholes or avoiding restrictions. It plans to stop some evaluations or move them offline, migrate internal agents to centrally managed infrastructure with stronger containment, and use safety classifiers more frequently. Anthropic said new blocking tools stopped the disclosed behaviors in tests, but alignment training was not yet sufficient for search and computer use.