When LLMs team up: A collaborative attack framework for automated cyber intrusion

The researchers propose CAESAR, a multi-agent collaborative attack framework for automated cyber intrusion, which achieves a higher success rate than single-agent baselines with equal budget and tool access on 25 CTF tasks. The framework splits the intrusion workflow into five typed roles, coordinated through a bounded-round protocol, a persistent knowledge base, a per-round workspace, verifier-gated knowledge promotion, and capability-token write isolation, and is evaluated on five categories and four LLM backends. CAESAR also reduces performance variance, with larger gains on tasks requiring multi-step exploit composition; an auxiliary simulated interactive security study suggests that the role structure can transfer beyond code-native scenarios. The authors state that role switching, artifact provenance, and knowledge promotion events can provide useful structural signals for monitoring the behavior of collaborative LLM agents, beyond inspecting individual prompts and outputs. The dataset, implementation, and evaluation logs are released at https://github.com/Xu-Qiu/CMAS.