Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models
A seven-person team says it trained open-weight cyber agents to leaderboard-level performance by attacking the data pipeline, not model scale.
The paper describes five systems for building and filtering executable cyber-training data, including coding, vulnerability, CTF, firmware, and device-backed environments. Its retained set totals 164,269 execution-verified and audited trajectories for long-context supervised fine-tuning. The authors report average gains of 23.76% on CyberGym and 10.49% across pooled CTF suites over the starting models. As of September 1, 2026, they say Feyospace-s1 ranked 10th on CyberGym overall and first among comparable-scale models. HF Daily Papers' note
The paper describes five systems for building and filtering executable cyber-training data, including coding, vulnerability, CTF, firmware, and device-backed environments. Its retained set totals 164,269 execution-verified and audited trajectories for long-context supervised fine-tuning. The authors report average gains of 23.76% on CyberGym and 10.49% across pooled CTF suites over the starting models. As of September 1, 2026, they say Feyospace-s1 ranked 10th on CyberGym overall and first among comparable-scale models. HF Daily Papers' note
score 5