Chinese artificial intelligence models are rapidly closing the cybersecurity gap with leading US systems, raising new questions about Washington’s efforts to restrict access to its most advanced technology.

GLM 5.2, a new open weight model from China’s Zhipu AI, matched or surpassed Anthropic models in some software vulnerability tests, according to security researchers.

Semgrep found that GLM 5.2 scored 39% on one benchmark designed to detect access control vulnerabilities, compared with 32% for Claude Code. The Chinese model also performed at a fraction of the cost.

The results do not mean GLM 5.2 has matched Anthropic or OpenAI across every task. It continues to trail leading US models in some broader reasoning and general capability tests.

However, its performance suggests that access to advanced cybersecurity capabilities is spreading beyond a small group of closed US systems.

GLM 5.2 is open weight, meaning users can download the model, operate it on their own hardware and modify it without relying on Zhipu’s infrastructure.

That makes the model attractive to businesses seeking lower costs and greater control over their data. It also creates security concerns because malicious actors can operate the system privately and remove safeguards.

The model has quickly gained adoption through platforms that provide access to multiple AI systems, reflecting growing demand for lower cost Chinese alternatives.

China’s 360 Security Technology has also introduced Tulongfeng, an AI system that the company claims can match Anthropic’s Mythos in finding software vulnerabilities. 

The company said the system had identified thousands of potential security flaws, although those claims have not been independently verified.

The advances come as the US government places new restrictions on domestic frontier models.

OpenAI limited the initial release of GPT 5.6 to a small group of approved partners following a government request for additional security reviews.

Anthropic also suspended access to its Fable 5 and Mythos 5 models after US officials imposed restrictions over concerns that foreign users could bypass their safeguards. Limited access to Mythos 5 has since been restored for some trusted US organizations.

Critics argue that the approach could undermine American AI companies while encouraging businesses to adopt Chinese open weight models that remain widely available.

US companies face uncertainty over whether access to powerful domestic systems could be restricted or withdrawn. Chinese alternatives, by comparison, can often be downloaded and operated independently.

That difference could accelerate adoption even if Chinese models remain weaker in some areas.

The shift also complicates US export controls. Restrictions on advanced chips and models were designed to preserve the country’s technological advantage, but techniques such as model distillation have allowed Chinese developers to learn from the output of leading American systems.