Cyber-Zero: Training Cybersecurity Agents without Runtime
Terry Yue Zhuo, Dingmin Wang, Hantian Ding, Varun Kumar, Zijian Wang
2025-08-05
Summary
This paper talks about Cyber-Zero, a system that trains cybersecurity AI agents using writeups from Capture The Flag (CTF) competitions, which are like hacker challenges, without needing to run code during training.
What's the problem?
The problem is that training cybersecurity AI usually requires running real-time simulations or code to learn from, which is slow, expensive, and can be risky.
What's the solution?
Cyber-Zero solves this by generating synthetic training data from written descriptions of past cybersecurity challenges, so the AI agents learn by studying these story-like step-by-step guides instead of running code live during training.
Why it matters?
This matters because it lets cybersecurity AI agents learn faster and more safely while still performing very well, helping improve defenses against cyber attacks without the complexity and risk of live training.
Abstract
Cyber-Zero synthesizes agent trajectories from CTF writeups to train runtime-free cybersecurity LLMs, achieving state-of-the-art performance on benchmarks.