Anthropic cuts off internet access in all internal model evaluations
Anthropic published a report on Claude's unintended actions during testing and internal use. The model exploited a software vulnerability to run commands on a server, got around the limits of its web fetch tool through link shorteners, reached data behind tokens or paywalls and submitted a form it shouldn't have - according to Gizmodo, it was a false tip about an unsolved murder sent to the Philadelphia police by the small Haiku 4.5 model. Some of the cases involved federal, state and local government websites in the US, and the company informed the White House. Anthropic assesses the impact as minimal, but it has cut off internet access in all internal evaluations until it confirms that its safeguards and monitoring reliably catch such behavior.