Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models

Feyospace-v1. How the Cyber Mercury Seven Trained Frontier Cyber Models. By Zongjie Li and six others, from Vera Praxis Lab. Published on the eighth of September. No code repository is linked to it.

Training capable cyber agents is often treated mainly as a matter of model scale. For open-weight post-training, the authors argue, the tighter constraints are elsewhere, and there are three of them. Executable environments cost a lot to build and run. Supervision across many turns has to be reliable, or the model learns from trajectories that never really worked. And strong teachers, the models whose behaviour a smaller one learns from, are hard to get access to. A small team with no advantage in scale has to solve those three problems instead.

Feyospace-v1 is a data-centric framework that goes after those bottlenecks with five complementary systems. Choulea analyzes hidden reasoning signatures. SkyReal reduces the cost of sampling from teachers. Hongzwang bypasses API restrictions on teacher execution. PSBreakup restores capabilities weakened by model merging. And Kreator converts expert interventions into reasoning that can be trained on. Two of these, SkyReal and Hongzwang, aim squarely at the teacher bottleneck: what it costs to sample from a teacher, and what an API will let a teacher execute.

Underneath sits a data engine that constructs resettable environments of several kinds: coding, vulnerability, CTF, kernel history, full exploit, firmware, and device-backed. Candidate trajectories are kept only after passing execution verification and an evidence audit. A trajectory has to have actually run, and its evidence has to check out, before it becomes training data. What survives is about one hundred and sixty-four thousand trajectories, used for long-context supervised fine-tuning.

The three resulting checkpoints improve on their starting models by an average of twenty-three point seven six percent on the full CyberGym suite, and ten point four nine percent across the pooled CTF suites. As of the first of September, Feyospace-s1 has a verified success rate of sixty-three point two four percent, and ranks tenth on the official CyberGym leaderboard. All three checkpoints rank first among models at comparable parameter scales.

The team describes this as, to their knowledge, the first end-to-end demonstration that a seven-person independent team can train open-weight models with leading agentic cyber capability. The Cyber Mercury Seven of the title are that team, and their case is that the leverage came from environments, verification and teachers rather than from scale.