Agent Against Agent: An Agentic System for Automatic Prompt Injection Red Teaming

The paper introduces PIMiner, an agentic system for automated prompt-injection red teaming. During training, it processes sequences of dataset–target-model pairs and builds a strategy library from scratch. At test time, the library transfers to previously unseen target LLMs without additional training, using only about 10 queries per target sample. On IPIArena, PIMiner reports attack success rates of 76.2% against Gemini-2.5-Pro, 61.9% against GPT-5.1, and 42.9% against Claude-Sonnet-4.5. On AgentDojo, it reports 86.7%, 53.3%, and 40.0%, respectively.