Training capable cyber agents is often framed as a problem of model scale, but open-weight post-training is constrained more directly by the cost of executable environments, reliable multi-turn supervision, and access to strong teachers. A new paper, Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models, presents a data-centric framework that tackles these bottlenecks through five complementary systems.
The systems are: Choulea, which analyzes hidden reasoning signatures; SkyReal, which reduces teacher-sampling cost; Hongzwang, which bypasses API restrictions on teacher execution; PSBreakup, which restores capabilities weakened by model merging; and Kreator, which converts expert interventions into trainable reasoning.
On the data side, the team built a data engine that constructs resettable environments spanning repository-level coding, vulnerability reproduction, CTFs, Linux kernel history, full exploit development, firmware, and device-backed systems. Candidate trajectories were retained only after execution verification and evidence auditing, yielding 164,269 audited trajectories for long-context supervised fine-tuning.
Without reinforcement learning, the resulting Feyospace checkpoints improve over their starting models by an average of 23.76% on the full CyberGym suite and 10.49% across pooled CTF suites. As of September 1, 2026, Feyospace-s1 achieves a verified success rate of 63.24% and ranks 10th on the official CyberGym leaderboard, while all three checkpoints rank first among models at comparable parameter scales.
The authors describe this as the first end-to-end demonstration that a seven-person independent team can train open-weight models with leading agentic cyber capability. An initial release, Feyospace-v1-small, is now available on Hugging Face, with more models and datasets planned.