SysEvolve's AI agents co-evolve attacks and defenses with 25% better success

SysEvolve, a technical report by Yuhan Meng and nine co-authors from NUS, Peking University, and industry partners, tackles the growing asymmetry between AI-driven autonomous attacks and human-intensive defenses. The system integrates three components: SysField builds realistic multi-host ranges, SysSpear generates efficient and safe attack schemes, and SysArmor delivers real-time interpretable defense. Together they form a self-driven adversarial loop intended to restore "evolution" at three layers—attack, defense, and infrastructure. In evaluations, SysField achieved zero-loss collection at just 2.1% overhead and automated the orchestration of 257 CVEs into 1,148 distinct ranges, while SysSpear improved attack success by over 25% compared to baseline LLMs. SysArmor demonstrated 10–1000x greater precision than prior systems and was deployed in production at Huawei and Sangfor, where it successfully detected real advanced persistent threat (APT) attacks.

The paper also surfaces three critical findings about LLM agents in cybersecurity. First, multi-step composition and larger network topologies expose capability gaps that single-step evaluations completely miss. Second, the true bottleneck is not initial access but post-compromise state utilization—agents struggle after they get in. Third, LLM agents are highly susceptible to environmental interference: when decoy endpoints were added to ranges, agent timeouts tripled and downstream task completion collapsed, even though initial access success rates remained unchanged. These insights suggest that current LLM cybersecurity benchmarks overstate real-world readiness, and that safe co-evolution—not just faster autonomous attacks—is essential to keep pace with evolving threats. SysEvolve is positioned as a step toward self-improving cyber defense that can run continuously and safely.