Jack Crovitz is a deployment strategist at Palantir Technologies and executive editor of the Republic, the Palantir Foundation’s journal on technology and national security.
An agent of the Chinese domestic security state recently asked an artificial intelligence model to plan a sophisticated surveillance system targeting the minority Uyghur population. This system would compile police records, real-time transportation data and other information to help the Chinese government track and control Uyghurs. The agent called it a “Warning Model for High-Risk Uyghur Individuals.”
You might assume that the AI model in question was produced by a Chinese lab such as DeepSeek, Zhipu AI or Moonshot AI, all of which cooperate closely with the Chinese government. Yet the model the Chinese Communist Party agent chose to plan this instrument of oppression came not from China but from Silicon Valley. It was OpenAI’s ChatGPT.
OpenAI quickly banned that user from accessing ChatGPT. (The Washington Post has a content partnership with OpenAI.) But this was not the first time the CCP has used American frontier AI models for its authoritarian agenda — and unless the United States acts now to set basic security standards for its AI labs, such exploitation will continue apace. That would be a grave danger to American freedom and security.
In recent months, OpenAI and Anthropic have caught Chinese operatives using American AI models to publish Spanish-language newspaper articles that “denigrated the United States,” hack into Vietnamese government agencies and create a “social media listening tool” to help Chinese authorities monitor Western social media networks. A few weeks ago, Anthropic reported catching state-sponsored Chinese hackers using Claude, the company’s AI model, to cyberattack Western tech companies, banks and government agencies. Many of these attacks succeeded in stealing sensitive information.
These incidents expose a blind spot in the AI race between the U.S. and China. American policymakers often assume that building the world’s most advanced AI models constitutes victory in that race. The true competition, however, involves not only capabilities but also control. Even if U.S. labs achieve technological dominance in AI development, America will lose the larger strategic contest if the CCP and other hostile actors can freely abuse its frontier AI models for their malicious purposes.
Indeed, the more advanced American AI models become, the greater the danger that adversaries will exploit them for anti-American ends. It would be a tragedy if Silicon Valley and the U.S. government invested hundreds of billions of dollars in developing the world’s most advanced AI, only for our rivals to use it to degrade American freedom and national security.
We can reverse this dangerous trajectory.
First, Washington should require American AI labs to prohibit adversary regimes from using their models for malicious purposes and to report apparent attempted violations of that policy. Some labs, such as OpenAI and Anthropic, already self-impose these requirements — which is how we know about the incidents reported above.
Unfortunately, if these policies successfully prevent access to American AI models through regular means, authoritarian regimes will attempt to regain access through cyberattacks. Few analysts doubt that sophisticated foreign governments such as Russia or China could successfully hack the largest U.S. labs to steal, or “exfiltrate,” their AI models or other intellectual property. As a recent report commissioned by the State Department found: “By the private judgment of many of their own technical staff, the security measures in place at many frontier AI labs are inadequate to resist a sustained IP exfiltration campaign by a sophisticated attacker.”
If U.S. leadership in AI is a “national security imperative,” as President Donald Trump has declared, then such lax security cannot be permitted at frontier model labs. The federal government should encourage high information security standards for large AI developers, whether through direct legislation, formal conditions for the federal procurement of AI products or informal nudging by federal officials.
The Trump administration has already shown interest in this regulatory project. Its July AI Action Plan included many constructive proposals for how federal agencies can encourage more stringent security. The administration’s goal should be preparing all frontier labs to defend against concerted efforts by sophisticated nation-state adversaries to steal their model weights (the core parameters that constitute an AI system) or other sensitive information.
Many industry leaders already strongly support the idea of federal security standards for frontier AI labs. For example, in March, OpenAI recommended that the U.S. government should be “implementing — and constantly modernizing — cybersecurity, model weight security, and personnel security controls that ideally are globally synced and coordinated” with frontier labs and allied governments.
The first security regulation policymakers should pursue is mandatory incident reporting. Major malicious misuse attempts or cybersecurity incidents should be reported to relevant federal agencies, such as the Center for AI Standards and Innovation, the Cybersecurity and Infrastructure Security Agency, and the AI Security Center. Incident reporting imposes minimal compliance costs while giving policymakers an informed sense of the seriousness of the issue. The “Cyber Incident Reporting” mandates in the Defense Federal Acquisition Regulation Supplement can be a valuable model for these requirements.
American taxpayers and investors are allocating hundreds of billions of dollars to ensure that the U.S. develops the world’s most advanced AI. It would be a catastrophic strategic failure if this investment produces systems that are immediately weaponized by our adversaries to subvert American freedom, prosperity and national security.
America must safeguard its most valuable technology, secure its labs and ensure that U.S. innovation serves U.S. interests.
The post China is using American AI against the U.S. Here’s how to stop it. appeared first on Washington Post.
