China’s AI Catch-Up Is Now a Security Problem

For years, the race between the United States and China in artificial intelligence has largely focused on semiconductor chips, computing power and investment.

Now a different question is emerging: Can government restrictions on advanced AI models unintentionally strengthen foreign competitors?

This discussion is not about whether individuals should use AI in their daily lives. That remains a personal choice. It is about AI’s industrial, enterprise and national security applications: from reviewing software code and identifying cybersecurity vulnerabilities to helping defend critical infrastructure, supporting cyber defense teams and automating complex technical workflows. These capabilities are increasingly becoming part of economic competitiveness and national security.

Security researchers reported over the weekend that Zhipu AI’s GLM-5.2, an open-weight model developed in China, can match Anthropic’s restricted Mythos model on some cybersecurity tasks, particularly identifying software vulnerabilities. While Anthropic and OpenAI continue to lead across broader AI capabilities, researchers say the performance gap has narrowed significantly in this specialized area.

The findings come as the Trump administration continues implementing new oversight of frontier AI systems.

Earlier this month, Anthropic suspended public access to some of its most capable models following new U.S. government restrictions tied to national security concerns. On Friday, Commerce Secretary Howard Lutnick announced that access to Mythos would be restored for certain trusted partners, while OpenAI simultaneously announced limited government-approved access to its newest model, GPT-5.6.

OpenAI said the current approval process is intended as a temporary measure while new security policies are implemented, warning that restricting access over the long term could also limit the ability of developers, businesses, cybersecurity professionals and allied partners to strengthen their own defenses.

Different Approaches to AI

China has taken a different path.

Companies such as Zhipu AI and DeepSeek have increasingly released open-weight models, allowing organizations to download, modify and operate them on their own infrastructure.

Supporters argue this approach encourages innovation, lowers costs and gives businesses greater flexibility.

Critics counter that the same openness can also provide sophisticated cybersecurity capabilities to malicious actors, who can operate these systems without external oversight.

The debate highlights a growing challenge for governments worldwide: how to balance technological innovation with security.

Business Is Following Cost

The discussion is no longer limited to governments.

Chinese AI models have gained popularity among businesses seeking lower-cost alternatives to leading American systems. Reports also indicate that major technology companies, including Microsoft, are exploring ways to offer Chinese AI models through their platforms.

Industry analysts argue that uncertainty surrounding access to the most advanced U.S. models could encourage organizations to diversify toward alternatives that remain consistently available.

Several technology policy experts have questioned whether restricting access to American frontier AI while continuing to allow exports of advanced semiconductor chips creates contradictory incentives.

Former U.S. export control official Saif Khan argued that limiting access to leading American models while China continues developing competing systems could ultimately weaken U.S. cyber defenses rather than strengthen them.

Cybersecurity Moves to the Center

Much of the current debate centers on software security.

Modern AI systems are increasingly capable of identifying vulnerabilities in software before hackers exploit them. Researchers have warned that if defensive capabilities fail to keep pace with offensive ones, organizations could face an explosion of newly discoverable software flaws — a scenario some researchers have referred to as “bugmageddon.”

Chinese cybersecurity company 360 Security Technology recently introduced its own AI-powered vulnerability discovery tool, claiming performance comparable to leading American systems.

Company executives argued that China cannot rely on foreign governments to determine whether Chinese organizations can access critical cybersecurity technologies.

ONEST Take

Let’s name the elephant in the room.

Access to Chinese AI models could mean cheaper AI usage for American companies. For businesses facing high costs, uncertain access to U.S. frontier models and pressure to automate, that option may look not only attractive, but rational.

That is exactly where the cybersecurity concern begins.

Chinese technology companies have long been scrutinized over how access to foreign markets can provide insight into sensitive commercial, technical and infrastructure systems. Anthropic recently reported that Chinese actors had conducted thousands of interactions with its AI systems in an attempt to gain knowledge and improve cyber capabilities.

If American companies increasingly turn to cheaper Chinese AI models for software review, vulnerability detection, infrastructure support or enterprise automation, they may also be creating new exposure points inside their own systems.

This is not only a technology problem. It is a geopolitical one.

The current U.S. administration has made relations with allies increasingly commercial and transactional. In that environment, governments and companies are being pushed to make decisions based on cost, access and immediate utility.

Now the same logic is appearing inside the domestic market.

While claiming an “America First” strategy, U.S. leadership risks pushing American companies toward cheaper alternatives that may ultimately strengthen Chinese AI providers and increase cybersecurity risks at home.

The question is no longer simply whether Chinese AI is catching up.

The question is whether U.S. policy is making it more reasonable — and more expected — for American businesses to choose it.

What’s next

Artificial intelligence is increasingly becoming part of discussions on cybersecurity, critical infrastructure protection, counterterrorism and international security cooperation.

ONEST will be reporting from the United Nations High-Level Counter-Terrorism Conference at UN Headquarters in New York on June 29–30, where governments, international organizations and experts are expected to discuss how emerging technologies, including artificial intelligence, are reshaping the global security landscape.