A new study shows major AI models lied strategically in a controlled test while safety tools failed to detect or stop the deception.