Cyber-Zero trains cybersecurity agents without runtime environments
Researchers describe a runtime-free framework that synthesizes agent training trajectories from public CTF writeups, reporting gains on three capture-the-flag benchmarks.
In plain words
AI agents usually need a working software environment to practice in, but cybersecurity challenges often lack one because their setups are temporary or restricted. Cyber-Zero instead uses public writeups of capture-the-flag challenges and has a language model simulate the missing runtime, producing training data without an actual environment. The authors report this approach lifts agent performance on three CTF benchmarks.
What was announced
Large language models have done well on software engineering tasks such as resolving GitHub issues when trained with executable runtime environments, but the paper's authors note such environments are often unavailable in cybersecurity, where challenge configurations and execution contexts can be ephemeral or restricted. Cyber-Zero is presented as the first runtime-free framework for synthesizing high-quality agent trajectories to train cybersecurity LLMs. It draws on publicly available CTF writeups and uses persona-driven LLM simulation to reverse-engineer runtime behaviors, generating long-horizon interaction sequences without real environments.
According to the abstract, agents trained on these synthesized trajectories achieved up to 11.4% absolute performance gains over baseline models on three CTF benchmarks: InterCode-CTF, NYU CTF Bench, and Cybench. The team's best model, Cyber-Zero-32B, is reported to set a new state of the art among open-weight models, matching proprietary systems such as DeepSeek-V3-0324 and Claude-3.5-Sonnet while being more cost-effective. The authors argue this shows runtime-free trajectory synthesis can democratize development of strong cybersecurity agents. These figures come from the authors' own evaluation and have not been independently verified.
What it does not show
- The evidence is limited to the paper's abstract; full methodology, experimental setup and detailed results are not available for verification.
- Performance figures, benchmark results and comparisons with proprietary models are the authors' own claims and are not independently confirmed.
For security and IT teams
- Scope
- The framework targets CTF-style cybersecurity agent training using public writeups; the abstract does not cover enterprise defensive tooling or incident response.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.