Reproduce “SecPO: Principled Adversarial Training for Prompt Injection Security” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sizhe Chen, Yibo Peng, Jaewon Chang, Chawin Sitawarin, David Wagner
SecPO trains assistants to resist indirect prompt injections in explicitly labeled untrusted data. It uses adaptive attack training and sets its reference to the undefended model’s clean-input response. On Qwen3.6-27B, it reports 8.8% AgentDojo adaptive attack success rate, versus 78.8% for Meta-SecAlign, with up to 800 candidates per pair. This does not establish general protection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.