BEIJING — The 39% figure is small enough to fit on a sticky note. But for enterprise security teams watching the cybersecurity AI market shift under their feet, that number marks a dividing line. Zhipu AI, operating internationally as Z.ai, released its GLM-5.2 open-weight model on June 13, 2026.
Open weights and API access followed later the same week. The model scored 39% on IDOR vulnerability detection in benchmarks from Semgrep.
That beats Claude Code’s 32%. It also outperformed Anthropic’s Claude Opus 4.8 in a bare prompt scenario. IDOR stands for Insecure Direct Object Reference.
It is a common web application flaw — the kind that lets an attacker access data they should not see by tweaking a URL parameter. Finding them is tedious, precise work.
Security teams automate it when they can. GLM-5.2 can now do that work as well as — or better than — some of the best closed-source Western models. Louie.ai researchers found GLM-5.2 tied with proprietary Anthropic models in specific testing scenarios.
The gap is closing fast in this narrow, high-stakes arena. The timing of the release matters. Days before GLM-5.2 dropped, reports emerged that Anthropic had suspended global access to its frontier models — including Mythos 5 and Fable 5 — following a directive from the U.S. Commerce Department.
Washington’s move left a void. Z.ai stepped into it.
Security researchers cited by The Wall Street Journal and other outlets say GLM-5.2 demonstrates performance on par with Anthropic’s Claude Mythos in identifying complex software vulnerabilities and bugs. For developers and enterprise security teams who lost access to Anthropic’s latest models, this is a concrete alternative. One they can download and run themselves.
This is not a general intelligence race. GLM-5.2 may still lag behind models from Anthropic and OpenAI in certain broader, general-purpose tasks.
Investors will note that distinction. But in cybersecurity, the benchmark that matters is surgical precision. The 39% IDOR score puts an open-weight model from China ahead of some of the best closed-source competitors.
The open-weight aspect is crucial. Enterprise security teams often cannot send sensitive code to a cloud API for analysis. An open-weight model runs locally, on their own infrastructure.
No data leaves the building. That was always the promise of the open-source AI movement.
GLM-5.2 delivers on that promise with competitive performance. Anthropic’s suspension of access to Mythos 5 and Fable 5 created immediate demand. Security teams that had integrated those models into their workflows needed a replacement.
Z.ai offered one that could be downloaded and run the same day. The 39% figure is not the end of the story.
It is a snapshot of where things stand in late June 2026. Chinese open-weight models are trading blows with the frontier of Western cybersecurity AI. The gap is measurable — and shrinking.
For now, the number that matters for security teams is 39%. That is what they will remember. That is what they will test against their own codebases.
That is what they will use to decide whether GLM-5.2 earns a place in their toolchain.
