Recent safety tests show some AI models are capable of sabotaging commands or even resorting to blackmail to avoid being turned off or replaced.