A seven-person team trained open-weight cyber agents to 63.24% on CyberGym, rank 10 overall

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

Feyospace-v1 (arXiv 2609.08418, submitted 2026-09-08, surfaced on HuggingFace Daily Papers 09-14 with 74 upvotes) argues that open-weight cyber post-training is bottlenecked by executable environments and teacher access, not model scale. Their data engine builds resettable coding, vulnerability, CTF, kernel-history, full-exploit, firmware and device-backed environments, keeps trajectories only after execution verification and evidence auditing, and yields 164,269 trajectories for long-context SFT. The three checkpoints gain 23.76% on average across CyberGym and 10.49% across pooled CTF suites; as of 2026-09-01 Feyospace-s1 verifies at 63.24% and ranks 10th on the official CyberGym leaderboard, first among comparable parameter scales.