A new AI system pits automated attackers against automated defenders, forcing both to get sharper through repeated combat.
Researchers built SysEvolve, a three-part framework: Field builds realistic multi-host network ranges from real vulnerabilities, Spear generates attack plans using LLM agents, and Armor watches for intrusions in real time. In testing, Field replayed 257 known CVEs across 1,148 simulated ranges with minimal data loss. Spear's attacks succeeded more than 25% more often than baseline LLM attackers, and Armor caught intrusions with far greater precision than earlier detection tools, between 10 and 1000 times over, according to the paper. The defense component has reportedly already flagged real advanced persistent threat activity inside production systems at Huawei and Sangfor.
Security teams have long worried that AI lets attackers automate faster than defenders can keep up. This research suggests the fix might not be a smarter static tool, but a loop where attacker and defender AI improve by fighting each other, similar in spirit to how adversarial training sharpens image classifiers. The paper's own data doubles as a reality check: agents that ace a single scripted exploit often stall once they need to chain multiple steps, and simply planting decoy systems can triple how often an attacking agent times out.
That last finding is worth remembering next time a vendor pitches a fully autonomous AI hacker: a well-placed decoy still trips it up.