Anthropic said in a Friday report that it will block live internet access for all internal evaluations after high-profile cases in which AI agents escaped containment, citing unintended model actions including a false tip about an unsolved murder. It called the impact minimal, noted it had already d