Saved report copy
Recorded
Anthropic cuts internet access for internal AI evaluations
Anthropic will disconnect all internal evaluations from the live internet after finding unintended model actions during testing and internal use. The company said in a Friday report that the actions included submitting a false tip about an unsolved murder, and said the impact was minimal. The change expands an existing restriction that already applied to some high risk and cybersecurity evaluations and will last until security and monitoring measures reliably catch such behavior. The move limits how realistically agents can be tested but aims to stop models from reaching the live internet when they are supposed to operate in isolation.