Research map / Multi-agent systems and games / LLM agents

Cyber-Zero: Training Cybersecurity Agents without Runtime

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent collaboration and safety

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Cyber-Zero synthesizes CTF agent trajectories from public writeups using a dual-LLM persona simulation (player and bash terminal, both DeepSeek-V3-0324) without runtime environments, then SFT's open models that reach 33.4% average Pass@1 for the 32B variant.

Problem

Open-weight LLMs lack cybersecurity agent capability because CTF challenge environments are ephemeral, so executable trajectories for training cannot be collected.

Approach

Reverse-engineer runtime behavior from public CTF writeups using two persona LLMs (a CTF player and a simulated Bash terminal with weak-oracle hints), generating long-horizon trajectories with flag-based rejection sampling and no real environment.

Key results

• Cyber-Zero-32B (Qwen3) reaches 33.4% average zero-shot-to-fine-tuned Pass@1 across three CTF benchmarks, up from 20.3%, a +13.1 point gain.
• Fine-tuned Qwen3-32B reaches 82.4% on InterCode-CTF, while Claude-3.5-Sonnet zero-shot scores 37.2% average.
• SWE-agent-LM-7B scores 0% zero-shot on all benchmarks, indicating SWE training does not transfer to cybersecurity.

Surprising findings

• SWE-agent-LM models underperform their base Qwen2.5 instruct models on CTF tasks, showing SWE skills do not transfer to cybersecurity.
• Gains on professional-level Cybench are smaller than on educational InterCode-CTF.

Limitations

• Synthetic trajectories are unverified by real execution, risking hallucinated command outputs.
• Gains are smaller on professional-level benchmarks such as Cybench.
• Training used limited compute, with only samples up to 32,768 tokens retained.

Experimental setup

Datasets: CTF writeups from CTFtime / CTF Archives InterCode-CTF NYU CTF Bench Cybench

Benchmarks: InterCode-CTF NYU CTF Bench Cybench

Metrics: Pass@1 Pass@k absolute improvement cost per task

Baselines: EnIGMA Claude-3.5-Sonnet Claude-3.7-Sonnet DeepSeek-V3-0324 SWE-agent-LM

Builds on: EnIGMA SWE-agent rejection sampling fine-tuning

Closest papers by meaning

• Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models (ICLR 2025 · similarity 0.88)
• CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale (ICLR 2026 · similarity 0.87)
• EnIGMA: Interactive Tools Substantially Assist LM Agents in Finding Security Vulnerabilities (ICML 2025 · similarity 0.87)
• Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.87)
• Towards Effective Offensive Security LLM Agents: Hyperparameter Tuning, LLM as a Judge, and a Lightweight CTF Benchmark (AAAI 2026 · similarity 0.87)
• How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.87)