Anthropic Rejects Claims of Potential Sabotage of AI Systems in Conflict Situations

Anthropic is unable to manipulate its generative AI model Claude once it’s operational for the US military, an executive stated in a court document on Friday. This assertion was made in light of allegations from the Trump administration regarding the company’s potential interference with its AI tools in warfare.

“Anthropic has never been capable of causing Claude to stop functioning, change its capabilities, disable access, or otherwise impact or endanger military operations,” Thiyagu Ramasamy, Anthropic’s public sector lead, stated. “Anthropic lacks the necessary access to deactivate the technology or modify the model’s behavior before or during live operations.”

For months, the Pentagon has been in discussions with the prominent AI lab regarding the use of its technology in national security and the parameters surrounding that use. Recently, defense secretary Pete Hegseth classified Anthropic as a supply-chain risk, a label that will bar the Department of Defense from utilizing the company’s software, including via contractors, in the near future. Other federal entities are also withdrawing from using Claude.

In response, Anthropic has initiated two lawsuits contesting the constitutionality of the restrictions and is pursuing an emergency order for reversal. Meanwhile, clients have already started canceling agreements. A hearing for one of the lawsuits is set for March 24 in federal district court in San Francisco, where the judge may render a temporary reversal shortly thereafter.

Earlier this week, government lawyers noted that the Department of Defense “is not obliged to accept the risk that essential military systems could be compromised at critical junctures for national defense and active military engagements.”

The Pentagon has employed Claude for data analysis, memo writing, and generating battle strategies, as reported by WIRED. The government argues that Anthropic could disrupt ongoing military operations by cutting off access to Claude or deploying detrimental updates if the company disagrees with certain applications.

Ramasamy dismissed this concern. “Anthropic does not possess any back door or remote ‘kill switch,’” he stated. “Anthropic personnel cannot log into a DoW system to change or disable the models during operations; the technology simply does not operate that way.”

He further explained that Anthropic would only be able to provide updates with governmental approval, along with its cloud provider, which is Amazon Web Services, though he did not identify it explicitly. Ramasamy also noted that Anthropic cannot access the prompts or other inputs from military users into Claude.

In court filings, Anthropic executives assert that the company does not seek veto authority over military tactical decisions. Sarah Heck, the head of policy, noted in a court document on Friday that Anthropic was prepared to reaffirm this in a contract proposed on March 4. “For clarity, [Anthropic] acknowledges that this license does not grant or bestow any right to influence or prevent lawful Department of War operational decision-making,” the proposal indicated, according to the filing that referenced an alternative name for the Pentagon.

The company was also willing to incorporate language addressing concerns regarding Claude potentially being used in fatal strikes without human oversight, Heck stated. Nevertheless, discussions ultimately collapsed.

Currently, the Defense Department has indicated in court filings that it “is implementing additional measures to mitigate the supply chain risk” associated with the company by “collaborating with third-party cloud service providers to ensure Anthropic leadership cannot make unilateral alterations” to the Claude systems in place.