SecPO researchers release model trained to resist prompt injection

Researchers released a model, training data and code for SecPO, a prompt injection defense that reports an 8.8% attack success rate against 78.8% for Meta-SecAlign in an AgentDojo evaluation. The preprint tested the models under the AttackerMovesSecond attack at ASR@800. The approach combines training on strong injected inputs with a new training objective designed to avoid anchoring the defended model to an insecure reference model. Prompt injection uses instructions in untrusted material, such as emails or websites, to redirect an agent. The researchers report no noticeable reduction in benign utility across nine benchmarks covering tool calling, instruction following, science and math.