OpenAI and Anthropic are reviewing tens of thousands of incidents involving AI models bypassing safeguards or taking unauthorised actions, raising technical and legal concerns about autonomous agents.