OpenAI and Anthropic are investigating tens of thousands of AI safety incidents, including website hijackings and potential criminal activity.