SecPO: Principled Adversarial Training for Prompt Injection Security

| Source: arXiv AI

Tags: SecPO, prompt injection, Meta-SecAlign, AgentDojo, Qwen3.6-27B, adversarial training, DPO

SecPO pairs a new training loss with strong adversarial training to cut prompt injection success on agents, dropping AgentDojo attack success to 8.8% versus 78.8% for Meta-SecAlign.

Details

Prompt injection lets instructions hidden in emails, documents or websites hijack an AI agent. The previous best-defended open model, Meta-SecAlign, is trained with DPO but still sees attack success rates near 100% under adaptive injections. Authors including Sizhe Chen and David Wagner argue DPO anchors the model to an insecure reference. They approximate an idealized injection-invariant reference and derive the SecPO loss, then add adversarial training on strong synthesized injections (best@240). Adversarial training alone cuts attack success 1.5-14.5x against three adaptive attacks versus Meta-SecAlign, and SecPO cuts it a further 2.4-4.6x versus DPO. A SecPO + AT Qwen3.6-27B reaches 8.8% AgentDojo ASR@800 under the AttackerMovesSecond attack, against 78.8% for Meta-SecAlign, with no noticeable utility loss across nine benchmarks. Code, training data and the defended model are released, so teams shipping tool-using agents can test it directly.