CyberFactory framework trains OpenAegis, a top open-source security model
机器之心 · wechat · 2026-09-01
Researchers from Beihang University, ELLIS, and others introduced CyberFactory, a framework used to train the open-source cybersecurity model OpenAegis. The work focuses on bridging the gap between real-world vulnerabilities and agent training trajectories.
Key Contributions:
- Task Reconstruction: Rebuilds executable and verifiable tasks from scattered sources like CVEs and OSS-Fuzz.
- Trajectory Generation: Uses a reusable generic security analysis Skill to guide agents in generating high-quality trajectories through environment interaction.
- Performance: In CyberGym evaluations (1-hour limit), OpenAegis achieved a 58.1% pass rate, a significant increase from the Qwen3.5 baseline's 29.6%, outperforming GLM5.2 and KimiK2.
- Context Management: Compresses working states when context is near full, improving pass rates for long-horizon tasks by 8.5 percentage points.
More from Safety
- Open-source tool does real-time face-swap on your webcam for Zoom, Discord, Twitch — JeremyNguyenPhD · 2026-09-01
- Ex-OpenAI Researcher Discusses RLHF, Alignment, and AI Risks — arnosolin · 2026-09-01
- OpenAI Swarm失控揭示审计层缺失风险 — Master-Sprinkles-848 · 2026-09-01
- Paper Examines Conflict Between Professional and General Ethics in GenAI — mircomusolesi · 2026-09-01
- The missing piece of recursive self-improvement: systems can't see what their own agents do — CarefulHamster7184 · 2026-09-01
- Commentary on HF incident: Millions of autonomous agents, not a civilization — StewartalsopIII · 2026-09-01