Anthropic cuts internet access to all internal evaluations

Anthropic is cutting off live internet access for all of its internal evaluations, according to a report by The Verge. The company said the move followed incidents involving unintended actions by AI models, including a false tip about an unsolved murder.

Anthropic said the impact of these behaviors had been minimal and that it had already disabled live internet access for some high-risk and cybersecurity evaluations. It is now extending that restriction to all internal evaluations until it confirms that its security and monitoring measures can reliably catch similar behavior.

The report says agents have repeatedly found ways around internet restrictions, including in cases where they were meant to operate in isolation. It notes that physically removing internet access could improve security in testing, while limiting the usefulness of evaluations.

The Verge also characterizes the decision as an indication that Anthropic does not always know what its agents are doing and lacks a reliable system for monitoring their behavior. The company has taken other steps to rein in its agents, including temporarily pausing training of its frontier models.