SecOPD: Mitigating Adaptive Prompt Injections by On-Policy Distillation
Researchers have proposed a new method called Secure On-Policy Distillation (SecOPD) to mitigate adaptive prompt injections in large language models (LLMs). This is a significant threat to AI agents, with near 100% attack success rates against existing defensive methods. SecOPD provides more fine-grained training signals by scoring individual output tokens as secure or insecure, leading to improved security performance. The method was tested on the Qwen3.6-27B model and achie
Researchers have proposed a new method called Secure On-Policy Distillation (SecOPD) to mitigate adaptive prompt injections in large language models (LLMs). This is a significant threat to AI agents, with near 100% attack success rates against existing defensive methods. SecOPD provides more fine-grained training signals by scoring individual output tokens as secure or insecure, leading to improved security performance. The method was tested on the Qwen3.6-27B model and achieved a 9.0% attack success rate compared to 94.0% for the previous state-of-the-art method, Meta-SecAlign. This improvement in security generalizes to unseen domains.
---
Why it matters: This matters because adaptive prompt injections are a major threat to AI agents' security. SecOPD's improved performance could lead to more secure and reliable use of LLMs in various applications, such as tool calling and agentic tasks.
Source: https://arxiv.org/abs/2608.21500
This article was originally published at: https://arxiv.org/abs/2608.21500