DoGNAVY-Exploit: Harnessing Parallel Open-Weight Agents for Stage-Aware Vulnerability Exploitation
| Source: arXiv AI
Tags: DoGNAVY-Exploit, GLM-5.3, Cybench, CVE-Bench, ExploitGym, Offensive Security
DoGNAVY-Exploit, a multi-agent harness for open-weight models, reports 97.5% Pass@1 on Cybench and solves 157 ExploitGym tasks with GLM-5.3 versus 130 for a public Claude Code baseline, narrowing the gap with frontier exploit agents.
Details
Frontier LLM agents can turn known vulnerabilities into working exploits but are costly and often access-restricted. Open-weight models struggle with long exploitation chains and applying security knowledge at the right stage. DoGNAVY-Exploit breaks exploitation into configurable milestones with stage-specific knowledge disclosure. Within each milestone a planner proposes attack routes, executors test them in parallel, and a controller accepts progress only with runtime evidence. A shared structured memory carries verified facts and failed routes across branches. In controlled benchmarks, GLM-5.2 reached 97.5% Pass@1 on Cybench and 97.5% zero-day / 100% one-day Pass@1 on CVE-Bench. On ExploitGym it solved 79 tasks with GLM-5.2 and 157 with GLM-5.3, against 39 and 130 for public Claude Code baselines.