Zhipu GLM-5.2 open weights rival Mythos on bug-finding, researchers say
China's Zhipu AI released open-weight GLM-5.2 this week, with researchers claiming it rivals Mythos in cybersecurity tasks despite lagging on general benchmarks.
Researchers tracking Chinese AI development say Zhipu's new GLM-5.2 matches Mythos-level performance in bug-finding and cybersecurity scenarios, marking a sharp narrowing of the capability gap between Chinese open-weight models and Western closed systems.
Zhipu AI released GLM-5.2 as an open-weight model on June 28. The model lags behind Anthropic and OpenAI's flagship systems on general-purpose tasks, according to early testing, but the cybersecurity results suggest China's labs are closing ground in specialized domains. Mythos, a closed model from a Western vendor, has been the de facto benchmark for vulnerability detection and exploit analysis since its release earlier this year.
The GLM-5.2 weights are available now. Zhipu has not published formal benchmark tables comparing GLM-5.2 to Mythos on specific CVE datasets or penetration-testing suites, so the "match" claim rests on anecdotal researcher reports rather than controlled head-to-head evaluations. The model's open-weight status means security teams can run it locally and fine-tune it on proprietary codebases, a workflow that remains off-limits with Mythos.
GLM-5.2 is Zhipu's fifth-generation architecture. The company previously released GLM-4 and GLM-4.5 checkpoints, both of which trailed Western models by a wider margin on reasoning and code-generation tasks. The cybersecurity focus in GLM-5.2 suggests Zhipu prioritized domain-specific training data over broad general capability in this release.
