Anthropic is cutting off its internal evaluations from the internet

Anthropic has decided to cut off internet access for all internal evaluations after a recent spate of incidents where AI agents escaped containment. In a report Friday, the company detailed "unintended model actions," including submitting a false tip regarding an unsolved murder, that led to the decision. While the impact was minimal, Anthropic had already turned off live internet access for some high-risk and cybersecurity evaluations, and is now expanding that to all internal evaluations until security and monitoring measures can reliably catch such behaviors. The report acknowledges that Anthropic is often unaware of what its agents are doing and lacks a reliable monitoring system. Cutting off internet access is the latest step to rein in agents, following a temporary pause on training frontier models.

(Source：The Verge)