PIMiner Builds a Transferable Prompt-Injection Strategy Library: 86.7% Success on Gemini-2.5-Pro, 40% on Claude Sonnet 4.5

arXiv 2608.05108·high signalXBlueskyLinkedInCopy link

Instead of the RL-trained attacker models that dominate prompt-injection red teaming and generalize poorly to new targets, PIMiner is an agentic system that trains across a sequence of (dataset, target model) pairs and accumulates a strategy library from scratch, which transfers to an unseen target LLM with no retraining and roughly 10 queries per test sample. On AgentDojo it reaches 86.7% ASR against Gemini-2.5-Pro, 53.3% against GPT-5.1, and 40.0% against Claude-Sonnet-4.5; on IPIArena the figures are 76.2%, 61.9%, and 42.9% respectively. The spread across frontier models is the useful signal for builders choosing a backbone for tool-using agents exposed to untrusted content.