Chimera-RL: An End-to-End Autonomous Red-Teaming Framework for LLM Applications

Large Language Model (LLM)–integrated enterprise applications introduce a dynamic and evolving attack surface, where vulnerabilities arise not from static code defects but from semantic misalignment, unsafe tool invocation, and multi-turn contextual reasoning. Existing security practices—including manual red teaming, prompt fuzzing, heuristic guardrails, and adversarial prompt libraries—remain limited in scalability and depth: they are labor-intensive, predominantly short-horizon, and rarely uncover multi-vector or multi-step exploit chains, while also lacking reproducible forensic evidence and systematic mitigation support. These limitations complicate the rigorous validation and hardening of complex LLM–tool–Retrieval Augmented Generation (RAG) systems. This paper presentsChimera-RL, an end-to-end autonomous red-teaming framework that operationalizes established reinforcement learning and attack-graph concepts within a unified system tailored to LLM-integrated applications. Rather than introducing a new learning algorithm, Chimera-RL combines curriculum-driven Hierarchical Reinforcement Learning (HRL), parameterized macro-actions, and mission-graph–guided exploration to enable long-horizon, campaign-level vulnerability discovery. The framework further integrates cryptographically hash-chained forensic logging and a template-driven mitigation playbook generator, providing verifiable attack traces and actionable defensive guidance.Experimental results demonstrate that Chimera-RL converges faster and more stably than standard deep RL baselines, includingProximal Policy Optimization (PPO)andAdvantage Actor–Critic (A2C), while discovering shorter and lower-cost exploit paths. When the generated mitigation playbooks are deployed, the attack success rate is reduced by nearly half, the median steps-to-compromise increase by a factor of seven, and the attacker’s operational cost more than doubles. These results indicate that Chimera-RL provides a practical and scalable system-level approach for automated red teaming and security hardening of LLM-integrated systems.