Researchers found a single, untouched 70 billion parameter language model can defend networks of wildly different sizes without retraining, something today's reinforcement learning defenders can't manage.
The paper, posted to arXiv this week, tests whether frozen large language models can replace reinforcement learning (RL) agents in automated cyber defense. The researchers built a two-layer system: a planner picks which subnet to protect, and an executor picks the specific defensive moves within it. They ran six models, from 3 billion to 70 billion parameters, including two built specifically for cybersecurity, against an automated attacker modeled on the MITRE ATT&CK framework inside the Cyberwheel simulation. Three configurations were compared: RL handling both planning and execution, LLM planning with RL execution, and LLM handling both.
Swapping only the planner for an LLM barely helped as networks grew larger. But handing execution to a capable LLM made a real difference: the 70B model kept successful lateral movement by attackers to about 1 percent of steps and attacker impact near zero, using identical weights across small, medium, and large networks. The RL baseline, by contrast, needed retraining for every network size.
That is the actual selling point here, not another AI stops hackers headline. A defender that does not need to relearn the network every time it grows is closer to something you would trust in production, though a university testbed is still a long way from a real enterprise network.