US Pentagon struggles to control military AI risks as US-China AI race accelerates.
America is entering a new AI arms race, but its own government is struggling to keep up. The US military, intelligence agencies and defense contractors are rapidly adopting artificial intelligence for cybersecurity, weapons, intelligence and national defense—yet Washington is proving unable to establish consistent controls.
The confusion became apparent in mid-July when the US Air Force ordered military contractors to remove Anthropic's AI products from weapons and control systems by September 1, with companies risking their Pentagon business if they failed to comply. About a month later, the Pentagon reversed course, telling the same contractors they did not have to immediately remove Anthropic products from Air Force systems and networks. The new guidance could still be reversed depending on litigation outcomes.
Defense Secretary Pete Hegseth has been in public dispute with Anthropic CEO Dario Amodei over restrictions the company wanted on how its AI could be used by the military. The Pentagon viewed these restrictions as unacceptable. Anthropic had sought limits on two particularly sensitive applications: the company refused to provide products to the military without assurances they would not be used for domestic surveillance of Americans or for fully autonomous weapons operating without human control. The original 25 pages of restrictions were eventually reduced to these two conditions.
The dispute took on new urgency when Anthropic introduced Mythos, a powerful AI system designed to find weaknesses in computer software—technology that could identify and potentially exploit vulnerabilities in highly protected networks, making it especially important for cyberwarfare. The National Security Agency, however, continued testing Anthropic's Mythos technology, arguing that without it, finding and fixing vulnerabilities in US military systems would become much harder. The NSA has been allowed to use experimental versions of Mythos despite broader Pentagon restrictions, testing it for both defensive and offensive cyber operations—using the technology to test American military networks while searching for weaknesses in networks belonging to China, Russia, North Korea and Iran.
One recently departed senior US official described stopping Mythos completely as similar to "unilateral disarmament," since the technology could significantly change the cyberwarfare balance. Banning a powerful AI tool could leave America weaker against rivals.
AI is already being deployed across America's national security system far beyond chatbots. US experts estimate that China could be only six months or less behind America in developing the most advanced AI models. That matters because advanced AI could dramatically amplify existing threats. Chinese hackers have already penetrated American communications and infrastructure networks in attacks known as Volt Typhoon and Salt Typhoon. Former Cybersecurity and Infrastructure Security Agency chief Jen Easterly said that what has changed is AI's ability to dramatically increase the speed and scale of those attacks.
The central debate is between safety and speed. Supporters of faster AI development fear that stricter rules could cost America its roughly six-month lead over China, while critics worry that moving too quickly could allow dangerous systems to escape human control. This debate is now shaping US national security policy.
Washington has struggled to create consistent rules for the most powerful AI models. Earlier this month, the Trump administration brought executives from major AI companies including OpenAI and Anthropic into discussions about reviewing new models for risks involving biological weapons, classified US systems and communications networks linked to nuclear weapons. Yet the policy has a major loophole: open AI models are exempt from the review process. These models can have their code customized by users, potentially modified for dangerous purposes—a concern raised by technology executives briefed on the policy.
This is especially problematic because many Chinese AI companies are developing open models. If America places heavy restrictions on its own private AI companies while Chinese companies continue developing customizable models, the policy could create a competitive disadvantage for the United States.
The deeper problem is that America may need the same AI companies it is trying to control. The Pentagon wants control over military AI applications, while companies such as Anthropic want safeguards around uses they consider too dangerous. This conflict has become increasingly difficult to manage as AI capabilities advance.
America's AI policy has been inconsistent on other fronts as well. The Trump administration has repeatedly changed its position on whether advanced Nvidia chips—crucial because powerful AI models require enormous computing power to train and operate—should be sold to China. This has created a balancing act for the White House: restricting American AI companies too heavily could push foreign countries toward Chinese AI models, while weak regulation could increase the risk of dangerous AI systems being used against the United States. A senior fellow at the Council on Foreign Relations and former Biden administration official described the administration's export-control policy as "incoherent."
The situation resembles the early nuclear arms race, though AI is fundamentally different. During the Cold War, the United States had a nuclear weapons advantage and feared Soviet catch-up. CIA Director John Ratcliffe recently said the comparison between AI and "digital nuclear weapons" was not misplaced. But AI cannot be controlled like nuclear weapons. Nuclear technology was largely controlled by governments, while AI is being developed primarily by private companies and spreads quickly across borders, making arms control far more complicated.
AI is also changing far faster than traditional military technology. Some of the most powerful AI models are updated weekly, and researchers often discover new capabilities only after systems are already being used. This speed and unpredictability make traditional national security planning much harder.
China may be behind America in advanced semiconductor production even if it is catching up in AI models. US experts estimate that China could be months behind in developing leading AI models but several years behind in producing the advanced chips needed to power them. That semiconductor gap is one reason the US is using export controls as part of its AI strategy, hoping that limiting China's access to advanced chips will make it harder for Chinese companies to build the computing infrastructure needed for frontier AI development.