New court filing reveals Pentagon told Anthropic the two sides were nearly aligned — a week after Trump declared the relationship kaput

Anthropic submitted 2 sworn declarations to a California national tribunal precocious Friday afternoon, pushing backmost connected the Pentagon’s assertion that the AI institution poses an “unacceptable hazard to nationalist security” and arguing that the government’s lawsuit relies connected method misunderstandings and claims that were ne'er really raised during the months of negotiations that preceded the dispute.

The declarations were filed alongside Anthropic’s reply little successful its suit against the Department of Defense and travel up of a proceeding this coming Tuesday, March 24, earlier Judge Rita Lin successful San Francisco.

The quality traces backmost to precocious February, erstwhile President Trump and Defense Secretary Pete Hegseth publically declared they were cutting ties with Anthropic aft the institution refused to let unrestricted subject usage of its AI technology.

The 2 radical who submitted the declarations are Sarah Heck, Anthropic’s Head of Policy, and Thiyagu Ramasamy, the company’s Head of Public Sector.

Heck is simply a erstwhile National Security Council authoritative who worked astatine the White House nether the Obama medication earlier moving to Stripe and past Anthropic, wherever she runs the company’s authorities relationships and argumentation work. She was personally contiguous astatine the February 24 gathering wherever CEO Dario Amodei sat down with Defense Secretary Hegseth and the Pentagon’s Under Secretary Emil Michael.

In her declaration, Heck calls retired what she describes arsenic a cardinal falsehood successful the government’s filings: that Anthropic demanded immoderate benignant of support relation implicit subject operations. That claim, she says, simply isn’t true. “At nary clip during Anthropic’s negotiations with the Department did I oregon immoderate different Anthropic worker authorities that the institution wanted that benignant of role,” she wrote. She besides points retired that the Pentagon’s interest astir Anthropic perchance disabling oregon altering its exertion mid-operation was ne'er raised during negotiations. Instead, she says, it appeared for the archetypal clip successful the government’s tribunal filings, which gave Anthropic nary accidental to respond.

Another item successful Heck’s declaration definite to gully attraction is that connected March 4 — the time aft the Pentagon formally finalized its supply-chain hazard designation against Anthropic — Under Secretary Michael emailed Amodei to accidental the 2 sides were “very close” connected the 2 issues the authorities present cites arsenic grounds that Anthropic is simply a nationalist information threat: its positions connected autonomous weapons and wide surveillance of Americans.

The email, which Heck attaches arsenic an grounds to her declaration, is notable for different reason. Two days later, aft Amodei said the institution was having “productive conversations” with the Pentagon, Michael posted connected X that “there is nary progressive Department of War dialog with Anthropic.” A week aft that, helium told CNBC determination was “no chance” of renewed talks. Heck’s constituent appears to be: If Anthropic’s stance connected those 2 issues is what makes it a nationalist information threat, wherefore was the Pentagon’s ain authoritative saying the 2 sides were astir aligned connected precisely those issues conscionable days aft the designation was finalized?

Ramasamy brings a antithetic benignant of expertise to the case. Before joining Anthropic successful 2025, helium spent six years astatine Amazon Web Services managing AI deployments for authorities customers, including classified environments. At Anthropic, he’s credited with gathering the squad that brought its Claude models into nationalist information and defence settings, including the $200 cardinal contract with the Pentagon announced past summer.

His declaration takes connected the government’s assertion that Anthropic could theoretically interfere with subject operations by disabling the exertion oregon different altering however it behaves, which Ramasamy says isn’t technically possible. Per his telling, erstwhile Claude is deployed wrong a government-secured, “air-gapped” strategy operated by a third-party contractor, Anthropic has nary entree to it; determination is nary distant termination switch, nary backdoor, and nary mechanics to propulsion unauthorized updates. Any benignant of “operational veto” is simply a fiction, helium suggests, explaining that a alteration to the exemplary would necessitate the Pentagon’s explicit support and enactment to install. Anthropic, helium says, can’t adjacent spot what authorities users are typing into the system, fto unsocial extract that data.

Ramasamy besides disputes the government’s assertion that Anthropic’s hiring of overseas nationals makes the institution a information risk. He notes that Anthropic employees person undergone U.S. authorities information clearance vetting — the aforesaid inheritance cheque process required for entree to classified information, adding successful his declaration that “to my knowledge,” Anthropic is the lone AI institution wherever cleared unit really built the AI models designed to tally successful classified environments.

Anthropic’s suit argues that the supply-chain hazard designation — the archetypal ever applied to an American institution — amounts to authorities retaliation for the company’s publically stated views connected AI safety, successful usurpation of the First Amendment.

The government, successful a 40-page filing earlier this week, rejected that framing entirely, saying that Anthropic’s refusal to let each lawful subject uses of its exertion was a concern decision, not protected speech, and that the designation was a straightforward nationalist information call, not punishment for the company’s views.

The Pentagon lawsuit is not the lone ineligible substance connected Anthropic’s docket this Tuesday. Separately connected Friday, a national justice tentatively ruled that Reddit’s suit against the institution — which accuses Anthropic of scraping its contented without support to bid its AI — should beryllium sent backmost to authorities court, wherever Reddit primitively filed it past June. A proceeding to finalize that determination is besides scheduled for Tuesday.