Answers are grounded in this paper and cite their sources.
In recent years, Autonomous Cyber Defense (ACD) methods driven by large language models (LLMs) have demonstrated promising performance in decision interpretability and environmental transferability, yet suffer from low decision efficiency and unstable decision-making. To mitigate these limitations, this work proposes a multi-agent hybrid LLM–expert system framework for ACD. The framework seamlessly integrates the determinism and efficiency of expert systems with the deep reasoning capabilities of LLMs: the expert system makes deterministic decisions based on experiential knowledge, whereas the LLM is responsible for complex decision tasks demanding deep exploration and long-chain reasoning. We further strengthen the LLM via chain-of-thought (CoT) guided prompting, historical memory, retrieval-augmented generation (RAG), output alignment and feedback learning. Experimental results demonstrate that incorporating the expert system improves both decision efficiency and defense performance. Evaluated on the CybORG CC4 environment, our approach achieves an average reward of − 98 over ten independent rounds, representing an approximate 25.5% improvement over the best-performing reinforcement learning (RL) baseline. Compared with the standalone LLM method, the average reward increases by roughly 96.2%, while the average per-step runtime is reduced by approximately 14.6%. To the best of our knowledge, this is the first work to systematically integrate LLMs and expert systems for ACD in the CybORG CC4 environment.
No takes yet. Share an insight, caveat, or question.
Qu et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: