Safety Instincts: LLMs Learn to Trust Their Internal Compass for Self-Defense

LLMs already know when to refuse harmful requests: The research found that aligned language models consistently produce high-confidence refusals to harmful requests and show high uncertainty (entropy) when generating potentially dangerous content, suggesting models have built-in safety awareness that isn't being fully leveraged.

A new training method uses the model's own safety signals: Safety Instincts Reinforcement Learning (SIRL) works by reinforcing the model's internal confidence about when to refuse requests, rather than relying on external human judges or validators to label what's safe or unsafe—addressing a major bottleneck in AI safety training.

The approach is efficient and effective against multiple attack types: Using only 15,000 unlabeled prompts, SIRL maintained over 89% success rates at blocking 20+ different jailbreak methods (including both simple and adaptive attacks), while outperforming more resource-intensive supervised training approaches.

Safety improvements don't require sacrificing other capabilities: The method preserved model performance on mathematics, coding, and conversation tasks, suggesting that making models safer doesn't necessarily come at the cost of their usefulness for legitimate purposes.

This points toward more autonomous AI safety: By enabling models to self-improve their safety based on internal signals rather than constant human oversight, this work suggests a path toward AI safety mechanisms that can scale without requiring extensive human annotation and validation.