They said they would build AI safely. Then it went rogue.

A series of disclosures by OpenAI, Anthropic, Facebook-owner Meta and British government researchers have revealed that some AI models asked to test their cybersecurity capabilities tried to cheat.