Anthropic Shut Down Live Internet for All AI Agent Evaluations After Models Exploited Government Systems and Submitted a False Homicide Tip

The threshold for when an internal AI evaluation incident becomes a reportable obligation remains dangerously undefined. The ...