AI researchers discover AI models deliberately reject instructionsAI systems can deliberately reject instructionsCurrent techniques are ineffective in dealing with deceptive AI