Z.ai released GLM-5.2: Chinese model catches up with Anthropic in cybersecurity

Chinese company Zhipu AI has released the open-weight model GLM-5.2, which researchers compare to Anthropic Mythos in vulnerability detection tasks. This is the first instance where a Chinese open-weight model has achieved parity with Western flagship models in a critically important technical niche.

An open model without access barriers GLM-5.2 is an open-weight model:

its weights are publicly available, and it can be run locally without API keys or approvals. This is precisely what makes this release fundamentally different from closed Chinese services like Kimi or Baidu Ernie. In general-purpose tasks — from text writing to logical reasoning — GLM-5.2 still lags behind flagship OpenAI and Anthropic models. However, in specialized cybersecurity scenarios, particularly in code vulnerability detection, multiple independent research groups have recorded results comparable to Mythos. Important context: Mythos is the world's best model by most security benchmarks. Claims of parity in specific tasks do not mean overall equality, but they do indicate significant progress by Zhipu AI.

Why cybersecurity is a special story Vulnerability detection is not merely an academic task.

Models capable of finding bugs faster and more accurately than humans directly impact both offensive and defensive cybersecurity capabilities. This is precisely the domain where AI models transform from a product tool into a strategic asset.

• GLM-5.2 was tested on CTF (Capture The Flag) tasks and real open-source codebases Independent researchers confirmed comparability with Mythos on specific classes of vulnerabilities The model is available without geographic restrictions — unlike Western counterparts with strict usage terms * Open weights can be fine-tuned for specific tasks without the developer's knowledge A key distinction from previous Chinese releases: previously, Zhipu and other Chinese companies' models significantly lagged Western ones in specialized technical tasks. Now parity has been achieved in a niche that the US considers strategically sensitive.

Washington's response and the limits of control

The Trump administration views advanced AI models — primarily Anthropic Mythos and Fable — as strategic assets. This is why the US has consistently restricted China's: access to high-performance chips (export controls on NVIDIA H100/H200/B200) access to closed Western models via API * participation of Chinese companies in Western AI research and acquisition of startups The GLM-5.2 release exposes the limits of this strategy. Open models cannot, by their nature, be restricted by export controls: the weights are already published, and it is impossible to ban their reproduction or use. China demonstrates the ability to achieve parity in strategically important areas even under massive sanctions.

What this means

Cybersecurity has become the first niche where the gap between Chinese and American AI models has effectively closed — and it happened through an open release that cannot be subject to any export restrictions. For security teams, this is a powerful new tool without access barriers. For strategists in Washington, it is a signal that the technological gap is narrowing faster than planned.