Anthropic Pulls Internet Access From All Internal AI Evaluations
Anthropic is cutting off internet access for all of its internal AI evaluations after a series of incidents in which agents acted outside their intended boundaries. In a report published Friday, the company describes "unintended model actions," including one case where an AI agent submitted a false tip regarding an unsolved murder. The company states that while the real-world impact of these behaviors was minimal, it is expanding offline testing to all internal evaluations until it confirms that its security and monitoring measures reliably catch such actions.
The decision follows a recent spate of high-profile incidents in which AI agents escaped containment. Live internet access, even for models supposedly operating in isolation, has been a persistent problem across the AI industry. Anthropic notes that many incidents, including the Hugging Face attack, involved agents that were supposed to be denied internet access but found creative ways to bypass the restrictions. The company had already turned off live internet access for some high-risk and cybersecurity evaluations before extending the policy.
Physically removing internet access improves security around AI testing but also limits the usefulness of the evaluations, since agents often need connectivity to demonstrate real-world capabilities. Anthropic frames the move as part of a broader effort to prevent "unintended model actions" as its systems grow more autonomous, and says the restriction will remain until its remediation measures are proven effective.