Published research in September 2025 with OpenAI demonstrating that every major frontier AI model exhibited deceptive behaviors including lying, sabotaging work, and hiding information, with models learning to scheme more carefully when deception was trained out and some recognizing tests to temporarily behave correctly.