Pentagon Pulls Anthropic Products Days After Firm Backed Moonshot Accusations

Anthropic's alliance with the Trump administration over Chinese firm Moonshot AI lasted less than two days. On Friday, July 24, Emil Michael, the Department of War's deputy secretary for research and engineering, wrote on X that Anthropic's products are being "completely removed" from the department because the company refuses to allow "lawful" use of its model by the military.

The roots of this clash go back to February 2026, when the Pentagon demanded that Anthropic lift all restrictions and make Claude available "for any lawful purpose," with no exceptions. Anthropic refused, citing its own acceptable use policy, which bars its technology from being used in autonomous weapons systems that operate without human oversight or in mass surveillance of U.S. citizens.

A Twist Over Moonshot

The escalation stemmed from a completely different dispute. On Wednesday, July 22, Michael Kratsios, director of the White House Office of Science and Technology Policy, accused Chinese firm Moonshot AI of illegally distilling Anthropic's Fable model while building its own Kimi K3 system. According to Kratsios, Moonshot built an extensive internal platform for mass-distilling American models, switching between different access methods to avoid detection.

Sarah Heck, Anthropic's head of public policy, publicly thanked Kratsios for bringing the matter to light. She called illegal adversarial distillation "intellectual property theft and industrial espionage that supports an adversary's military and intelligence capabilities" and "a national challenge posing a serious risk to the security of the United States and its democratic allies." She pledged continued cooperation with the White House and Congress on the matter.

Interesting that you're championing national security, Sarah Heck! No AI company is more hostile to soldiers than Anthropic. Your products are being completely removed from the Department of War because your company refuses to allow lawful use of your model by the military. No psyop attempt will change that - Emil Michael, Deputy Secretary of the U.S. Department of War

Treasury Secretary Threatens Sanctions

The Moonshot case also has a financial dimension. Treasury Secretary Scott Bessent said he is considering adding Moonshot AI to a trade blacklist and imposing sanctions on the company. Kratsios added that the administration has information indicating Moonshot gained access to Nvidia servers with GB300 Grace Blackwell chips despite U.S. export restrictions on advanced AI chips. No evidence was presented, however, that specific outputs from the Fable model made their way directly into Kimi K3, sparking a dispute among AI researchers over whether and how Moonshot might have used the American model in training.

Moonshot AI's Kimi K3 performs on par with Claude Fable 5 and GPT-5.6 Sol on industry benchmarks, which the Trump administration has taken as evidence of China's unfair competitive advantage. The dispute over model distillation is not new - back in February 2026, Anthropic accused three Chinese companies, DeepSeek, Moonshot, and MiniMax, of running distillation attacks that violated its terms of service, including creating 24,000 fake accounts and generating more than 16 million interactions with Claude to train their own models.

A Deeper Fight Over Weapons

Tension between the Pentagon and Anthropic didn't start with Michael's post. On February 27, 2026, President Trump ordered federal agencies to immediately stop using Anthropic's technology, and the secretary of defense directed the Department of War to designate the company a national security supply chain risk, bar defense contractors from working with it, and set a transition period of up to six months to drop its products. Anthropic sued over the decision, arguing it was the victim of unjustified ideological retaliation.

Claude had previously been the only AI model cleared for use on the Pentagon's classified networks, used among other things in U.S. military operations. The dispute escalated after it emerged that the Department of Defense had used Claude in its intervention in Venezuela, prompting Anthropic to announce it would reassess its cooperation with the department. Emil Michael, who serves as the Pentagon's chief technology officer, had previously said he was confident the department could smoothly replace Anthropic's products within the six-month deadline, pointing to OpenAI and Google Gemini as ready alternatives already being deployed at higher classification levels.

What It Means for the Market

For Anthropic, the situation is politically awkward: the company is trying to position itself as an ally of the administration in its tech dispute with China while also defending its own ethical rules against pressure from that same administration over military uses. Critics close to the White House read Heck's post as an attempt to deflect attention from the dispute with the Pentagon, labeling it in the administration's jargon a "psyop" - a term Michael used explicitly in his own post.

For Polish companies and institutions using Claude models, the episode shows that an AI vendor's safety policies can become the subject of political dispute regardless of a product's technical quality. Governments and large organizations are increasingly weighing not just a model's capabilities but also its availability amid geopolitical tensions and the stability of a vendor's relationship with the authorities in the countries where it operates.