Agent Against Agent: An Agentic System for Automatic Prompt Injection Red Teaming
PIMiner transfers a learned prompt-injection attack strategy library to unseen LLM agents with about 10 target queries per sample.
The paper positions this as an alternative to RL-based red-teaming methods that the authors say generalize poorly across target models. PIMiner trains over dataset-and-model pairs, building its strategy library from scratch, then applies it at test time without additional training on the new target. The reported attack success rates include 61.9% against GPT-5.1 on IPIArena and 53.3% on AgentDojo, with higher rates against Gemini-2.5-Pro and lower rates against Claude-Sonnet-4.5. Code is listed as available. HF Daily Papers' note
The paper positions this as an alternative to RL-based red-teaming methods that the authors say generalize poorly across target models. PIMiner trains over dataset-and-model pairs, building its strategy library from scratch, then applies it at test time without additional training on the new target. The reported attack success rates include 61.9% against GPT-5.1 on IPIArena and 53.3% on AgentDojo, with higher rates against Gemini-2.5-Pro and lower rates against Claude-Sonnet-4.5. Code is listed as available. HF Daily Papers' note
score 5