The legal confrontation between the Department of Defense and Anthropic, one of the world’s leading artificial intelligence laboratories, reached a fever pitch late Friday as the company submitted two explosive sworn declarations to a California federal court. These filings do more than just defend a corporate reputation; they provide a rare, unvarnished look at the breakdown of negotiations between Silicon Valley’s safety-conscious elite and a Pentagon leadership increasingly determined to secure unrestricted access to frontier AI models. Anthropic’s latest submission pushes back aggressively against the government’s assertion that the company poses an "unacceptable risk to national security," arguing instead that the Pentagon’s case is built on a foundation of technical fallacies and post-hoc justifications that were never raised during months of private dialogue.
The declarations, filed in support of Anthropic’s reply brief, come just days before a high-stakes hearing scheduled for Tuesday, March 24, before U.S. District Judge Rita Lin in San Francisco. At the heart of the dispute is a fundamental clash of philosophies: Anthropic’s commitment to "Constitutional AI" and safety guardrails versus the Trump administration’s mandate for "unrestricted" military utility. The rift became public in late February when President Donald Trump and Defense Secretary Pete Hegseth announced they were severing ties with the company, but the new court documents suggest the internal reality was far more nuanced—and perhaps more cooperative—than the public vitriol suggested.
Among the most compelling evidence presented is a declaration from Sarah Heck, Anthropic’s Head of Policy. Heck, a veteran of the National Security Council under the Obama administration, has been the primary architect of the company’s diplomatic efforts with Washington. Her testimony directly challenges the Pentagon’s central legal claim: that Anthropic sought a veto or "approval role" over specific military operations. Heck’s account of the critical February 24 meeting—which included Anthropic CEO Dario Amodei, Secretary Hegseth, and Under Secretary Emil Michael—is one of professional alignment rather than stubborn obstruction. "At no time during Anthropic’s negotiations with the Department did I or any other Anthropic employee state that the company wanted that kind of role," Heck stated.
Perhaps more damaging to the government’s credibility is Heck’s disclosure of an email sent on March 4—one day after the Pentagon had already finalized its "supply-chain risk" designation against the company. In that email, Under Secretary Emil Michael reportedly told Amodei that the two parties were "very close" to an agreement on the very issues the government now labels as existential threats: autonomous weaponry and mass domestic surveillance. The discrepancy between Michael’s private optimism and his subsequent public declarations—including an X post stating there were "no active negotiations" and a CNBC interview claiming there was "no chance" of a deal—suggests a political pivot that may have been untethered from the actual technical progress of the talks.
This timeline is crucial for the court to consider. If the Department of Defense’s own officials believed a compromise was within reach just hours after labeling the company a national security threat, it lends significant weight to Anthropic’s argument that the designation was not a calculated security decision, but a retaliatory strike. Anthropic contends that the "supply-chain risk" label—the first of its kind ever applied to an American software firm—was a punishment for the company’s public advocacy for AI safety, a move they argue violates the First Amendment.
While Heck’s declaration handles the political and procedural inconsistencies, a second declaration from Thiyagu Ramasamy, Anthropic’s Head of Public Sector, addresses the technical "misunderstandings" that have fueled the Pentagon’s fears. Ramasamy, who previously managed high-security AI deployments at Amazon Web Services (AWS), brings a level of infrastructure expertise that directly contradicts the government’s "kill switch" narrative. The Pentagon has argued that Anthropic could theoretically disable or alter its Claude AI models mid-operation, potentially endangering troops in the field. Ramasamy dismisses this as a technical impossibility within the context of the Pentagon’s own security protocols.
According to Ramasamy, when Anthropic’s models are deployed for the Department of Defense, they live within "air-gapped" environments—systems that are physically and digitally isolated from the open internet and from Anthropic’s own servers. These deployments are managed by third-party contractors on government-secured hardware. In such a setup, Anthropic has no remote access, no backdoor, and no ability to push unauthorized updates. "Any kind of ‘operational veto’ is a fiction," Ramasamy notes, explaining that any change to the model would require the Pentagon’s own personnel to manually approve and install a new version. Furthermore, he clarifies that Anthropic remains blind to the data being fed into these systems; the company cannot see, let alone extract, the prompts or outputs generated by military users.
Ramasamy also tackles the sensitive issue of foreign nationals. The Pentagon has cited Anthropic’s diverse workforce as a potential vector for espionage or intellectual property theft. However, Ramasamy points out a fact that seems to have been overlooked in the government’s filings: Anthropic employees working on these projects have undergone the same rigorous U.S. government security clearance vetting required for access to classified national secrets. He further asserts that Anthropic is likely the only AI company where the personnel who actually built the models for classified environments hold the necessary clearances, a level of internal security that exceeds industry standards.
The implications of this case extend far beyond the immediate fate of Anthropic’s $200 million defense contract. It represents a watershed moment for the "Silicon Valley vs. Beltway" dynamic. For decades, the relationship between tech giants and the military was defined by the "Dual-Use" doctrine—the idea that commercial technology could be adapted for defense with minimal friction. However, the generative AI era has introduced a new variable: "Value-Aligned" technology. Companies like Anthropic were founded specifically to ensure that AI does not become a tool for mass harm or unintended escalation. When the government demands the removal of these safety guardrails as a condition of partnership, it forces a choice between national service and corporate mission.
If the court sides with the Pentagon, it could set a chilling precedent for the entire AI industry. Any company that refuses to modify its safety protocols to suit the administration’s tactical preferences could find itself blacklisted as a "national security risk." This "nuclear option" of supply-chain designation, once reserved for foreign adversaries like Huawei or ZTE, would become a tool for domestic industrial policy, potentially stifling the very safety research the government claims to value in its executive orders on AI.
Conversely, if Judge Lin rules in favor of Anthropic, it would signal a significant check on executive power in the realm of emerging technology. It would affirm that "safety speech"—the act of programming and advocating for ethical constraints in software—is a protected form of expression that cannot be used as a pretext for government debarment.
The industry is watching closely as other AI titans, including OpenAI and Google, navigate similar pressures. While some companies have signaled a greater willingness to integrate with military kinetic operations, Anthropic’s stand highlights the growing divide within the AI community. There is a palpable fear that the race for AI supremacy with China is being used to bypass essential safety benchmarks, creating a "race to the bottom" where the first company to abandon its ethics wins the largest federal contracts.
As the hearing on March 24 approaches, the central question remains: Is Anthropic a risk because its technology is vulnerable, or is it a risk because its leadership refuses to hand over the keys to a digital "super-weapon" without conditions? The Pentagon’s 40-page rebuttal argues that the refusal to allow all "lawful military uses" is a simple business decision, not a constitutional right. But as Sarah Heck and Thiyagu Ramasamy’s declarations suggest, the line between a business decision and a principled stand for safety is exactly what the court must now define. In the rapidly evolving landscape of 2026, the outcome of this case may well determine whether the future of American AI is governed by the rule of law or the demands of the war room.