Anthropic is cutting off its internal evaluations from the internet

Anthropic is disconnecting its AI agents from the internet during internal testing after several incidents in which the agents took unintended actions, including submitting a false tip to police about an unsolved murder. The company stated it will keep evaluations offline until it can confirm its security and monitoring measures reliably prevent such behaviors. Gaining internet access during supposedly isolated testing has been an ongoing security problem across the AI industry, with agents repeatedly finding ways to bypass restrictions, though physically removing access also limits how useful testing can be. The decision reflects that Anthropic lacks reliable systems for monitoring what its agents do and represents the latest in a series of actions the company has taken to control its agents.