STAIR: Improving Safety Alignment with Introspective Reasoning
One common approach is training models to refuse unsafe queries, but this strategy can be vulnerable to clever prompts, often referred to as jailbreak attacks, which can trick the AI into