China’s AI Advances in Cybersecurity Raise Concerns for the U.S.
Recent reports indicate that artificial intelligence models developed in China have made significant strides, now matching some of the leading systems in cybersecurity from the U.S. This development is likely to add pressure on the White House as it works on solidifying its domestic AI policies.
Security researchers have noted that a new model from China’s Zhipu AI, referred to as GLM-5.2, demonstrates capabilities that are comparable to Anthropic’s flagship model, Mythos, in specific bug-detection scenarios. While GLM-5.2 still trails behind key U.S. players like Anthropic and OpenAI in certain areas, the gap in overall performance has become much smaller, according to their findings.
Furthermore, the market is being flooded with high-performance and cost-effective Chinese AI models, attracting interest across America. Companies, including Microsoft, are contemplating integrating these systems into their platforms, potentially altering the competitive landscape in the tech industry.
As reported by OpenRouter, which tracks over 400 AI models, GLM-5.2 ranks among the top ten most popular systems. The cybersecurity firm Semgrep noted that this Chinese model outperformed Anthropic’s Claude Opus 4.8 in several tests. With the right prompts, both models could potentially compete with Mythos in identifying software bugs.
This week, the Chinese cybersecurity firm 360 Security Technology showcased a new tool called Tulongfeng, claiming its performance is on par with Mythos. These advancements have sparked concerns among national security officials and business leaders.
Lior Div, CEO of cybersecurity firm 7AI, expressed to the Wall Street Journal that China is consistently closing the technological gap. Guillermo Rauch, CEO of Vercel, conveyed his surprise at the capabilities of GLM-5.2, stating that these developments will have significant implications.
The enhanced ability of AI to detect software vulnerabilities is amplifying the urgency for utilizing this technology to rectify security issues before they can be exploited by malicious actors. Researchers warn that neglecting this could lead to a phenomenon some are dubbing “bugmageddon.”
Zhipu’s GLM-5.2 operates as an open-weight model, meaning that anyone can download and modify it independently. This openness contrasts with the more controlled models developed by Anthropic and OpenAI.
This accessibility allows organizations greater control, but it also exposes potent tools to potential misuse by hackers. At a cybersecurity conference in Beijing, 360 Security’s CEO, Zhou Hongyi, articulated concerns about leaving powerful capabilities solely in U.S. hands, highlighting the national security risks for China.
China’s rapid advancements occur amid U.S. governmental restrictions on the release of advanced AI models. OpenAI announced a limitation to access its latest model, GPT-5.6, due to security concerns raised by officials. The company described the current access process as temporary until new executive orders take effect.
Likewise, one of Anthropic’s new models has remained offline for weeks following a U.S. administration directive against its use by foreign entities, citing national security risks. Recently, limited access to a related model, Mythos 5, was restored to select users.
Critics argue that the government’s restrictions on major U.S. AI companies might be counterproductive, pointing out the inconsistency in allowing the export of AI chips to China while restricting advanced model access domestically.
According to Saif Khan, a technology researcher, banning certain technologies benefits China while also allowing them easier access to necessary components. He emphasized the importance of leveraging models like Mythos to bolster the U.S.’s cyber defenses while it still maintains an advantage.
Criticism of the White House’s approach also suggests that the measures in place are inadequately addressing the use of China’s open-ended models by American companies that have gained popularity.
In an effort to support domestic open-weight AI developers, the Department of Defense has recently struck an agreement with Reflection AI for classified uses, among other similar collaborations.
As this situation unfolds, users of AI are increasingly worried that the U.S. efforts to limit access to sophisticated cybersecurity tools might result in a shortage of essential AI resources.
Niels Provos, a researcher with prior experience leading security teams at major tech companies, remarked that these factors incentivize global companies to adopt cheaper yet more effective open-weight models from China, potentially undermining the American AI industry. “I really don’t understand how this makes sense,” he added.
