AutoPI: Taxonomy-Guided Adaptive Red Teaming for Prompt Injection in LLM Agents
AutoPI: Taxonomy-Guided Adaptive Red Teaming for Prompt Injection in LLM Agents
Researchers Damianakis and Psounis propose AutoPI, an adaptive red teaming method that uses the prompt injection technique taxonomy PIAtlas to guide search and measure coverage. It generates injections for each sub-technique in the taxonomy, then rewrites failed injections based on feedback from an LLM judge; the goal is to find more diverse attack methods rather than only raising attack success rates. The method is black-box and training-free, can replace the fixed attack templates in existing benchmarks, and updates with the taxonomy. In evaluations on AgentDojo and IPI Arena against open-source and closed-source models, AutoPI found successful injections for Qwen3.6-35B in all six categories and over 90% of techniques; against Claude Haiku 5.5, the most robust model in the evaluation, its attack success rate was 20% to 50% higher than recent state-of-the-art adaptive red teaming methods, and its category-level coverage was 25% to 400% higher. The authors released the code and generated attacks to support reproducible, coverage-oriented evaluation.