PISmith: Reinforcement Learning-based Red Teaming for Prompt Injection Defenses

TL;DR

PISmith is a reinforcement learning framework that systematically evaluates prompt injection defenses by training attack models, revealing vulnerabilities in current defenses and outperforming baseline attack methods across multiple benchmarks.

Contribution

Introduces PISmith, an RL-based red-teaming framework that effectively assesses and exposes weaknesses in prompt injection defenses against adaptive attacks.

Findings

State-of-the-art defenses remain vulnerable to adaptive RL-based attacks.

PISmith outperforms 7 baseline attack methods across 13 benchmarks.

Effective in agentic settings against various LLMs.

Abstract

Prompt injection poses serious security risks to real-world LLM applications, particularly autonomous agents. Although many defenses have been proposed, their robustness against adaptive attacks remains insufficiently evaluated, potentially creating a false sense of security. In this work, we propose PISmith, a reinforcement learning (RL)-based red-teaming framework that systematically assesses existing prompt-injection defenses by training an attack LLM to optimize injected prompts in a practical black-box setting, where the attacker can only query the defended LLM and observe its outputs. We find that directly applying standard GRPO to attack strong defenses leads to sub-optimal performance due to extreme reward sparsity -- most generated injected prompts are blocked by the defense, causing the policy's entropy to collapse before discovering effective attack strategies, while the rare successes cannot be learned effectively. In response, we introduce adaptive entropy regularization and dynamic advantage weighting to sustain exploration and amplify learning from scarce successes. Extensive evaluation on 13 benchmarks demonstrates that state-of-the-art prompt injection defenses remain vulnerable to adaptive attacks. We also compare PISmith with 7 baselines across static, search-based, and RL-based attack categories, showing that PISmith consistently achieves the highest attack success rates. Furthermore, PISmith achieves strong performance in agentic settings on InjecAgent and AgentDojo against both open-source and closed-source LLMs (e.g., GPT-4o-mini and GPT-5-nano). Our code is available at https://github.com/albert-y1n/PISmith.