Court filing shows Pentagon said deal with Anthropic was close despite Trump ending talks

A new court filing has revealed that the Pentagon indicated it was close to reaching an agreement with Anthropic, even as public statements from U.S. officials suggested the relationship had already collapsed. The development emerged from sworn declarations submitted by Anthropic to a federal court in California late Friday, as the company challenges the government’s claims that it poses an “unacceptable risk to national security.”

The filings were submitted as part of Anthropic’s reply brief in its ongoing lawsuit against the Department of Defence. The case is scheduled for a hearing on Tuesday, March 24, before Judge Rita Lin in San Francisco. In its submission, Anthropic disputes the government’s position, arguing that the Pentagon’s claims rest on technical misunderstandings and on allegations that were never raised during months of prior negotiations between the two sides.

The conflict dates back to late February, when President Donald Trump and Defence Secretary Pete Hegseth publicly announced that the U.S. government would end its relationship with Anthropic. That decision followed the company’s refusal to permit unrestricted military use of its artificial intelligence systems.

Two senior Anthropic executives submitted sworn declarations in support of the company’s case: Sarah Heck, Head of Policy, and Thiyagu Ramasamy, Head of Public Sector. Heck, who previously served on the National Security Council during the Obama administration and later worked at Stripe, currently oversees Anthropic’s government relations and policy initiatives. She attended a key meeting on February 24 with Anthropic CEO Dario Amodei, Defence Secretary Hegseth, and Under Secretary of Defence Emil Michael.

In her declaration, Heck challenges what she describes as a fundamental inaccuracy in the government’s argument—namely, that Anthropic sought a role in approving or influencing military operations. She states clearly that no such demand was ever made during negotiations. According to Heck, neither she nor any other Anthropic representative suggested that the company wanted authority over military decision-making.

Heck also addresses the Pentagon’s claim that Anthropic could potentially interfere with operations by disabling or altering its technology mid-mission. She notes that this concern was never raised during the negotiation process and only appeared later in the government’s legal filings, leaving the company without an opportunity to respond at the time.

One of the most notable details in Heck’s declaration concerns an email sent on March 4, the day after the Pentagon finalised its supply-chain risk designation against Anthropic. In that message, Under Secretary Michael reportedly told CEO Dario Amodei that both parties were “very close” to agreement on two key issues that the government now cites as reasons for labelling the company a national security threat: its positions on autonomous weapons and the mass surveillance of Americans.

Heck included the email as evidence, highlighting what she suggests is a contradiction between internal communications and subsequent public statements. On March 5, Amodei described discussions with the Pentagon as “productive.” However, shortly thereafter, Michael stated publicly on X that there were no active negotiations between the Department of War and Anthropic. Within a week, he told CNBC that there was “no chance” talks would resume.

The timeline presented by Heck raises questions about the consistency of the government’s position. If Anthropic’s stance on autonomous weapons and surveillance was the basis for its designation as a national security risk, she implies, it is unclear why a senior Pentagon official would characterise the two sides as being close to agreement on those very issues immediately after the designation was made.

Ramasamy’s declaration focuses on technical aspects of the government’s claims. Before joining Anthropic in 2025, he spent six years at Amazon Web Services, where he managed AI deployments for government clients, including work in classified environments. At Anthropic, he helped build the team responsible for deploying the company’s Claude AI models within national security and defence contexts, including a $200 million Pentagon contract announced the previous summer.

In his statement, Ramasamy disputes the assertion that Anthropic could interfere with military operations by remotely disabling or modifying its technology. He explains that once the company’s AI systems are deployed in secure, air-gapped government environments managed by third-party contractors, Anthropic has no access to them. According to him, there is no remote “kill switch,” no hidden backdoor, and no mechanism for unauthorised updates. Any changes to the system would require explicit approval and action by the Pentagon itself.

Ramasamy also emphasises that Anthropic cannot view or access the data entered by government users into its systems, countering concerns about potential data exposure or misuse.

Additionally, he addresses the government’s argument that the company’s employment of foreign nationals poses a security risk. He states that Anthropic personnel involved in sensitive work have undergone U.S. government security clearance processes, the same vetting required for handling classified information. He further notes that, to his knowledge, Anthropic is unique among AI companies in having cleared personnel directly involved in building models intended for classified use.

Anthropic’s lawsuit contends that the Pentagon’s supply-chain risk designation—described as the first of its kind applied to a U.S. company—amounts to retaliation for the company’s publicly stated views on AI safety. The company argues that this action violates its First Amendment rights.

The government, however, rejects this characterisation. In a detailed 40-page filing submitted earlier in the week, officials argued that Anthropic’s refusal to support all lawful military applications of its technology constitutes a business decision rather than protected speech. They maintain that the designation was made on national security grounds and was not intended as punishment for the company’s positions or public statements.