RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution
RedEvoAgent is a black-box automated red-teaming agent for LLM-based tool-using systems that learns a compact, human-readable “attack skill” from prior attack trajectories instead of replaying fixed jailbreaks or retrieving whole past traces. Its key idea is to evolve that skill using tool-effectiveness profiling, attribution of which tool decisions actually helped, and a validation ratchet that keeps only performance-improving updates, which makes it more efficient and transferable than trajectory-retrieval baselines, though it still seems dependent on the quality of its available attack tools and validation setup.
Abstract
LLM-based agents are increasingly deployed in product-level execution harnesses, where jailbreaks can trigger harmful tool use and persistent state changes, creating greater risks than unsafe text generation alone. Existing automatic red-teaming methods often rely on fixed attacks, while recent agentic attackers coordinate multiple jailbreak tools and show stronger potential through trajectory-based retrieval. However, such retrieval can reuse misleading experiences due to retrieval bias and unclear tool credit, and full trajectories add context overhead while reducing interpretability. We propose RedEvoAgent, a black-box red-teaming agent that distills cross-case attack trajectories into a concise, human-readable attack skill. The attack skill adaptively evolves through tool-effectiveness profiling and Deciding-Tool Attribution for skill updates, and a validation ratchet that retains only updates improving validation performance. Experiments on multiple benchmarks, target models, and target execution harnesses show that RedEvoAgent outperforms fixed and agentic baselines, improves tool efficiency, and transfers across attacker models and target execution harnesses.