Court Filing Exposes Pentagon’s Covert AI Courtship with Anthropic, A Week After Trump Declared It ‘Kaput’

Late Friday afternoon, Anthropic presented two sworn affidavits to a federal court in California. These documents challenged the Pentagon’s declaration that the artificial intelligence firm constitutes an “intolerable threat to national security,” contending instead that the government’s legal arguments stem from technical misinterpretations and contentions not brought up throughout the prolonged negotiation period leading up to the disagreement.

These affidavits were lodged concurrently with Anthropic’s rebuttal brief in its legal action against the Department of Defense. Their submission precedes a forthcoming hearing scheduled for this Tuesday, March 24, to be presided over by Judge Rita Lin in San Francisco.

The disagreement originates from late February, a period when President Trump and Defense Secretary Pete Hegseth publicly announced their severance of relations with Anthropic. This decision followed the company’s refusal to permit unfettered military utilization of its AI solutions.

The two individuals who presented the affidavits are Sarah Heck, Anthropic’s Policy Chief, and Thiyagu Ramasamy, who leads the company’s Public Sector division.

Heck previously served as an official on the National Security Council, having been employed at the White House during the Obama administration. Subsequently, she transitioned to Stripe before joining Anthropic, where her responsibilities include overseeing the company’s governmental liaisons and policy initiatives. She was directly in attendance at the February 24 meeting, during which CEO Dario Amodei convened with Defense Secretary Hegseth and Emil Michael, the Pentagon’s Under Secretary.

Within her affidavit, Heck highlights what she characterizes as a fundamental inaccuracy within the government’s submissions: the assertion that Anthropic sought an authorization function concerning military undertakings. That contention, she affirms, is simply baseless. “During Anthropic’s discussions with the Department, neither I nor any other Anthropic staff member ever indicated the company desired such a capacity,” she penned. Furthermore, she underscores that the Pentagon’s apprehension regarding Anthropic potentially deactivating or modifying its technology during ongoing operations was never brought forth during the negotiation phase. Rather, she states, this concern surfaced initially in the government’s court documents, affording Anthropic no chance for rejoinder.

A further point in Heck’s affidavit likely to garner notice is that on March 4 — the day following the Pentagon’s official completion of its supply-chain risk classification against Anthropic — Under Secretary Michael conveyed via email to Amodei that both parties were “extremely near” agreement on the two matters the government currently references as proof that Anthropic constitutes a national security peril: its stances on self-governing armaments and widespread monitoring of U.S. citizens.

This electronic mail, which Heck includes as an attachment in her affidavit, moreover prompts inquiries regarding Michael’s subsequent public pronouncements. On the ensuing day, March 5, Amodei released a statement indicating the company had been engaged in “fruitful discussions” with the Pentagon. The day thereafter, Michael shared on X that “no current Department of War discussions are underway with Anthropic.” A week later, he informed CNBC that “no possibility” existed for recommenced negotiations. Heck’s argument seems to be: If Anthropic’s position on these two concerns renders it a danger to national security, how could the Pentagon’s own representative assert that both factions were almost in accord on those identical matters merely days subsequent to the designation’s finalization?

Ramasamy contributes a distinct form of specialized knowledge to the legal proceedings. Prior to his tenure at Anthropic starting in 2025, he dedicated six years at Amazon Web Services overseeing AI implementations for governmental clients, encompassing classified operational contexts. Within Anthropic, he is recognized for establishing the group responsible for integrating its Claude models into national security and defense applications, a feat that includes the $200 million agreement with the Pentagon publicized last summer.

His affidavit addresses the government’s assertion that Anthropic could hypothetically disrupt military endeavors by deactivating the technology or modifying its performance, which Ramasamy contends is not technically feasible. According to his account, once Claude is implemented within a government-protected, “air-gapped” infrastructure managed by an external contractor, Anthropic possesses no means of entry; there exists no distant deactivation switch, no clandestine entry point, nor any method to force unapproved revisions. Any notion of an “operational prohibition” is imaginary, he implies, clarifying that any modification to the model would necessitate the Pentagon’s explicit endorsement and undertaking for installation. Anthropic, he states, is even unable to view what government personnel input into the system, much less retrieve that information.

Furthermore, Ramasamy contests the government’s allegation that Anthropic’s recruitment of non-U.S. citizens renders the firm a security hazard. He points out that Anthropic’s workforce has successfully completed U.S. government security screening — the identical background verification procedure mandated for entry to confidential data. He further specifies in his affidavit that, “to my understanding,” Anthropic stands as the sole AI enterprise where vetted staff members genuinely constructed the AI models intended for deployment in secure, classified settings.

Anthropic’s legal complaint asserts that the supply-chain risk classification — the inaugural instance applied to a U.S. corporation — constitutes governmental reprisal for the company’s openly declared perspectives on AI security, thereby infringing upon the First Amendment.

The government, within a 40-page submission earlier in the week, repudiated that characterization completely. It contended that Anthropic’s unwillingness to permit all legitimate military applications of its technology represented a commercial judgment, distinct from safeguarded expression, and that the classification was a clear-cut national security determination, not a penalty for the firm’s opinions.

The Pentagon litigation is not the sole legal affair on Anthropic’s schedule for this Tuesday. Independently on Friday, a federal magistrate provisionally decreed that Reddit’s legal action against the corporation — alleging Anthropic harvested its content without authorization to educate its AI — ought to be remanded to state court, where Reddit initially initiated it last June. A session to conclude that determination is likewise slated for Tuesday.