Paper states SecPO+AT reduces Qwen3.6-27B's AgentDojo adaptive attack success rate to 8.8%
安全偏好优化与对抗训练,提升模型内生安全
Researchers from UC Berkeley, Anthropic and other institutions released the SecPO paper, proposing a prompt injection defense training method combining safety preference optimization with adversarial training. Under AgentDojo adaptive attacks, the method reduced attack success rate from Meta-SecAlign's 78.8% to 8.8%; on Qwen3.6-27B, the success rates of three adaptive attacks—Genetic, TAP and PISmith—dropped to 13.3%, 12.5% and 1.8% respectively, with normal task capability largely maintained. SecPO changes DPO's reference distribution from attacked inputs to model behavior under clean inputs, while AT dynamically searches for high-intensity attack samples during training based on PAIR. The paper notes that the defense presupposes the system can distinguish trusted instructions from untrusted data, and that of 140 successful DTaP attack cases, 75 executed malicious operations while completing the original task.
The article breaks down why traditional DPO struggles against adaptive prompt injection, and how SecPO adjusts the reference distribution and pairs it with adversarial training, and provides specific success rate comparisons.