Google AI Introduces Consistency Training for Safer Language Models Under Sycophantic and Jailbreak Style Prompts