SafeReview: Defending LLM-based review systems against adversarial hidden prompt attacks

The researchers propose a co-evolutionary adversarial training framework, SafeReview, to defend LLM-based review systems against manipulation by adversarial prompts hidden in submissions. It jointly trains a Generator to produce increasingly effective prompt injections and a Defender, trained via preference optimization, to keep reviews consistent between clean and attacked submissions. Experiments show that it resists adaptive prompt injection attacks better than static defenses, preserves paper rankings when under attack, and generalizes to different attacker architectures.