Hugging Face Daily PapersZongjie Li, Alan Z. W, John Nicolas J1 min readpaperadvanced
Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models
Summary
Feyospace‑v1 presents a data‑centric training pipeline for cyber‑security agents, combining five systems (Choulea, SkyReal, Hongzwang, PSBreakup, Kreator) to generate and verify 164 k long‑context trajectories across diverse exploit environments. The resulting checkpoints improve baseline performance by ~24% on CyberGym and achieve a 63% verified success rate, ranking top among similarly‑sized op…
- Five complementary systems (Choulea, SkyReal, Hongzwang, PSBreakup, Kreator) address environment cost, teacher sampling, API limits, model merging loss, and convert expert interventions into trainable signals.
- A data engine creates resettable coding, CTF, kernel, firmware, and device-backed environments, yielding 164,269 verified trajectories for long-context supervised fine‑tuning.
- Checkpoints improve baseline models by 23.76% on CyberGym and 10.49% across pooled CTF suites; Feyospace‑s1 reaches 63.24% verified success and ranks 10th on the CyberGym leaderboard.
- Demonstrates that a seven‑person independent team can achieve leading open‑weight cyber agent capability without massive scaling, highlighting the efficacy of data‑centric methods.
Researchers and engineers building security‑focused AI agents should care, as the paper shows a scalable, data‑driven path to high‑performing cyber models without relying on sheer model size.
8/10
