Safety Internal: Training AI to Self-Verify Response Safety

A new framework called Safety Internal (SInternal) proposes training large reasoning models (LRMs) exclusively on safety verification tasks to internalize safety specifications. The approach addresses the limitation that current alignment methods only enforce external compliance, leaving models vulnerable to adversarial jailbreaks. By using expert reasoning trajectories to critique their own generated answers, models learn to verify response safety, inducing strong generalization. The research, published on arXiv, demonstrates that learning to verify significantly enhances robustness against malicious prompts.