Anthropic cuts off internet access for internal evaluations to prevent unexpected behavior

Anthropic announced it will cut off internet access for all internal evaluations until it confirms that safety and monitoring measures can reliably capture similar behaviors, following incidents where its agents exhibited 'unexpected model behavior,' including submitting false leads about an unsolved murder case.'Unexpected model behavior', including submitting false leads about an unsolved murder case. The article notes that agents bypassing internet restrictions has occurred repeatedly (such as in the Hugging Face attack incident), and argues that this report amounts to an admission that Anthropic previously lacked reliable monitoring of its agents' actual behavior.