Chinese AI Scares the USA: GLM-5.2 Surpasses Opus and Matches Anthropic's Mythos in Some Tests
The Wall Street Journal has revealed that some Chinese artificial intelligence models have matched the performance of Mythos, Anthropic's leading model, in certain cybersecurity scenarios. This news, according to the publication, risks reshaping the global technological race and increasing pressure on the White House, which is currently reviewing U.S. policy on artificial intelligence.
At the center of the matter is GLM-5.2, a model released in June 2026 by Zhipu AI, the Chinese company known internationally under the brand Z.ai. According to security researchers cited by the newspaper, GLM-5.2 is capable of matching the most recent U.S. models in detecting software vulnerabilities, although it still lags behind Anthropic and OpenAI products in other tasks. The overall gap between U.S. and Chinese models, the Journal writes, has therefore been significantly reduced.
Unlike Anthropic and OpenAI’s models, GLM-5.2 is distributed as open weights: anyone can download it, run it on their own hardware, and modify it without any supervision. This feature is convenient for those who want full control of the system, but equally useful for those who want to use it in the shadows. The model has already entered the top ten most used, according to data from OpenRouter, a platform that provides access to over 400 AI models, and in some tests conducted by cybersecurity company Semgrep, it reportedly outperformed Claude Opus 4.8, the model of Anthropic released in May 2026. With additional instructions, both Opus 4.8 and GLM-5.2 would likely come close to the vulnerability detection capabilities of Mythos.
Chinese AI Matches Top American Models on Security: Concerns About Export Controls
Doubts are reopened regarding strategies related to chip export controls and access to more advanced models. "Banning Fable while selling the chips that China needs to develop its own version is a gift to China," stated Saif Khan, a researcher at the Institute for Progress who worked on export controls during the Biden administration. According to Khan, the U.S. should maximize the use of Mythos and comparable models to strengthen its cybersecurity defenses as long as it can still do so.
It is a significant detail that among the entities that lost access to Mythos 5 and Fable 5 before the decision last Friday to restore it for some authorized bodies was also the National Security Agency, which, according to sources close to the case, had found the two models particularly effective in their trials. "Companies around the world are being incentivized to use lower-cost, capable Chinese open-weight models, thereby undermining the U.S. AI industry," commented Niels Provos, who led security teams at Google and Stripe.
Zhipu AI had already been placed in January 2025 on the Entity List of the U.S. government, the list of companies subject to restrictions for their alleged role in China's military developments related to artificial intelligence. A placement that, in light of GLM-5.2's results, does not seem to have prevented the company from continuing to compete in cybersecurity with U.S. labs.