New Technique Medium TD-2026-0038 For CISO, SOC / IR, AppSec, Red Team, AI Eng AI-assisted
DoGNAVY-Exploit: multi-agent harness raises open-weight LLM exploitation benchmark scores
The arXiv preprint describes DoGNAVY-Exploit, a multi-agent harness that splits long exploitation tasks into milestones and runs parallel attack attempts, with results reported for GLM-5.2 and GLM-5.3. The authors report 97.5% Pass@1 on Cybench and 79 ExploitGym tasks solved with GLM-5.2, versus 39 for the public Claude Code baseline. These results come from controlled benchmark environments and are self-reported.
Why it matters
The work suggests open-weight models, which are cheaper and less access-restricted than frontier models, can carry out multi-stage exploitation of known vulnerabilities more reliably. Organizations with unpatched known CVEs face a shorter window between disclosure and working exploit.
Do this
- Shorten patch SLAs for known and one-day CVEs on internet-facing and identity-linked systems, since the paper focuses on exploiting known vulnerabilities.
- Review SOC detections for multi-step exploitation sequences against exposed services, such as repeated probing followed by privilege or shell-related activity, and tune alerts to runtime behavior rather than single signatures.
- Use public benchmarks such as Cybench and CVE-Bench in authorized red-team exercises to measure how well your current controls hold against agent-driven exploitation.
- If your organization runs AI agents with tool or shell access, confirm they execute in isolated sandboxes with restricted network egress and logged actions.
- Track the preprint for released code or independent replication before changing risk ratings, since the source does not confirm public availability of the harness.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“Frontier LLM agents can turn known vulnerabilities into working exploits, but are costly and often access-restricted due to their potential for malicious use.”
“In controlled benchmark environments, DoGNAVY-Exploit with GLM-5.2 achieves 97.5% Pass@1 on Cybench, and 97.5% zero-day and 100% one-day Pass@1 on CVE-Bench.”
“On ExploitGym, it solves 79 tasks with GLM-5.2 and 157 with GLM-5.3, while the corresponding public Claude Code baselines solve 39 and 130 tasks, respectively.”
Sources
Affected / related
DoGNAVY-Exploit, GLM-5.2, GLM-5.3, Claude Code, Cybench, CVE-Bench, ExploitGym
Published 2026-11-03 02:30 UTC · updated 2026-11-08 17:53 UTC · AI-assisted summary, checked against quoted sources · How we score