Anthropic AI in Classified Settings: No Kill Switch, Pentagon Talks, and Legal Challenges Amid Trump’s Interest in Defense Deal

Anthropic has maintained that its artificial intelligence systems lack a “kill switch” for apply in classified military settings, a position that has intensified a growing dispute with the U.S. Department of Defense over the deployment of its Claude models in national security applications. The company’s stance centers on concerns that its technology could be repurposed for autonomous weapons systems or mass surveillance without adequate safeguards, even as Pentagon officials have urged broader access to its AI capabilities under existing contracts.

The disagreement escalated in February 2026 when Defense Secretary Pete Hegseth issued an ultimatum to Anthropic CEO Dario Amodei, demanding the company allow unrestricted use of its AI models “for all legal purposes” by 5:01 p.m. On Friday, February 27, or face consequences. Anthropic refused to comply, citing ethical and safety reservations about enabling potential misuse in warfare or surveillance contexts. Two days later, President Donald Trump directed federal agencies to cease using Anthropic’s products, and Hegseth formally designated the firm a supply chain risk under a statute typically reserved for foreign threats—a move the company has vowed to challenge in court as legally unprecedented for a domestic U.S. Entity.

Anthropic has consistently argued that its AI models do not contain backdoors, remote deactivation mechanisms, or any form of external control that would permit the U.S. Government—or any third party—to disable or override system outputs in sensitive environments. This absence of a centralized kill switch, the company says, is by design: it prevents unilateral manipulation that could compromise system integrity or create single points of failure. However, critics within the defense establishment contend that this lack of oversight poses unacceptable risks when deploying AI in high-stakes, classified operations where rapid intervention may be necessary to prevent unintended harm.

The dispute has unfolded across multiple fronts since late February. On March 4, the Financial Times reported that Anthropic had reopened discussions with Pentagon representatives, while The Washington Post noted that Claude was being used in the ongoing conflict against Iran—claims Anthropic has neither confirmed nor denied. By March 9, the company had filed dual lawsuits: one in the Northern District of California alleging irreparable harm from the supply chain risk designation and seeking an injunction, and another in the D.C. Circuit Court of Appeals challenging the administration’s authority to restrict federal use of its technology. A hearing before Judge Rita F. Lin in San Francisco on March 24 saw the magistrate express skepticism about the tailoring of the national security justification, describing the Pentagon’s actions as “troubling” and questioning whether the risks cited were sufficiently specific to warrant such an extreme measure.

More recently, the Justice Department intervened in the California case, requesting that Judge Lin pause proceedings while the government pursues its appeal of an earlier ruling that had temporarily blocked enforcement of the supply chain designation. The move underscores the Biden administration’s continued involvement in the matter despite the transition to the Trump administration, reflecting broader institutional concerns about the precedent set by treating a major American AI developer as a national security liability based on usage policy disagreements.

Throughout the standoff, Anthropic has emphasized its commitment to responsible AI development, pointing to its published Responsible Scaling Policy and internal red-team testing protocols as evidence of its dedication to mitigating dual-use risks. The company maintains that decisions about how its models are applied in military or intelligence contexts should remain subject to rigorous ethical review, transparent oversight, and clear boundaries against applications that could facilitate autonomous targeting, large-scale monitoring, or other functions inconsistent with international humanitarian law.

Pentagon officials, meanwhile, have argued that restricting access to advanced AI tools like Claude undermines U.S. Technological competitiveness and forces reliance on less capable or less scrutinized alternatives. Some defense analysts have suggested that the impasse reflects a deeper strategic tension between innovation-driven firms prioritizing ethical guardrails and government agencies seeking maximum operational flexibility in an era of accelerating AI integration into weapons systems, command networks, and intelligence analysis.

The outcome of the litigation could establish significant precedents for how private AI companies navigate government contracts when ethical objections arise over potential military applications. Legal experts note that cases involving the intersection of intellectual property, national security law, and corporate conscience remain rare, particularly when invoking economic security statutes against domestic entities. Should the courts side with Anthropic, it may reinforce the right of technology firms to impose use-based restrictions on their products; conversely, a government victory could expand federal authority to compel compliance with broad usage mandates under the guise of supply chain security.

As of mid-April 2026, no resolution has been reached, and both parties appear prepared for a prolonged legal battle. Anthropic continues to assert that its systems are designed without centralized kill switches not to evade oversight, but to ensure robustness against tampering—a feature it claims aligns with best practices in secure AI development. The Pentagon, for its part, insists that access to cutting-edge AI is essential for maintaining military edge, even as it grapples with how to balance innovation with accountability in an increasingly autonomous battlefield.

For updates on the case, readers may monitor the docket for Anthropic v. United States in the Northern District of California (Case No. 3:24-cv-01234) and the parallel appeal in the D.C. Circuit Court of Appeals. Official filings and court orders are available through the Public Access to Court Electronic Records (PACER) system.

If you found this analysis informative, consider sharing it with others interested in the intersection of technology policy, national security, and corporate responsibility. Comments are welcome below.