SecPO adversarial training for prompt injection security
Original title: SecPO: Principled Adversarial Training for Prompt Injection Security
The SecPO loss combined with strong adversarial training reduces attack success rates by 1.5-14.5x versus prior methods in agentic workflows.
A new principled SecPO loss is introduced for prompt injection defense by approximating an injection-invariant oracle model. An efficient adversarial training pipeline synthesizes strong attacks for training. The resulting model on Qwen3.6-27B maintains utility while improving security over Meta-SecAlign.
From the original article
1 min read at arxiv.orgRecognized as a top threat to AI agents, prompt injection allows instructions from untrusted data (email, documents, websites) to manipulate the agent's operation. The existing best-defended LLM, Meta-SecAlign, is trained with the DPO loss but still suffers from attack success rates (ASRs) approaching 100% under adaptive prompt injections.
By Sizhe Chen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.