AI Battle: How We Pitted 4 Attack Agents Against MaxPatrol O2 and Who Won

Positive Technologies set up a cyber battle at the Standoff range, where four autonomous AI attackers targeted a corporate network defended by MaxPatrol O2. The attackers were: OpenCode + HexStrike AI (DeepSeek V4 Pro), OpenCode + NyxStrike (GLM 5.1), PentAGI (GLM 5.1), and Yasen (Anthropic Haiku 4.5 + Opus 4.6). All agents started by exploiting CVE-2024-11680 in ProjectSend to gain initial access, then escalated privileges, moved laterally, and executed different critical events: stealing confidential files, manipulating inventory data, extracting strategic documents, and obtaining SCADA operator credentials. Attack times ranged from 40 minutes to several hours. HexStrike AI achieved the highest success rate by completing all four scenarios. MaxPatrol O2 used an LLM-driven investigation agent to reconstruct kill chains, scoring based on sequence accuracy (SA) and structural similarity (SJS). Its best score was 0.73 out of 1. The experiment showed that AI agents can autonomously execute realistic attacks, but human-in-the-loop corrections were often needed. The defense system successfully identified most attack steps and captured credentials and host transitions.