Countdown pressure from the Department of Defense is increasing. Anthropic must comply with the Pentagon’s request by 5:01 p.m. ET or be designated a “supply chain risk,” a label typically applied to companies with suspected foreign influence. The background is simple. The Pentagon wants Anthropic’s Claude technology to be able to be deployed on sensitive networks “for any lawful purpose,” but the company has two red lines in place. The idea is that Claude will not be used in autonomous weapons or for mass surveillance of American citizens.
Anthropic said Thursday it has no intention of retracting its position.
“Threats do not change our position. In a clear conscience, we cannot agree to their demands.”
The Department of Defense emphasizes that both of the proposed use scenarios are not necessary and that there must be freedom to deploy licensed technologies within the bounds of safety and legality.
“This is a simple and prudent request from a common sense perspective, and will prevent threats to critical military operations and risks to our nation’s military.”
A tense meeting took place Tuesday at the Pentagon between Defense Secretary Pete Hegseth and Anthropic CEO Dario Amodei. Participants said the talks were friendly, but Pentagon officials threatened to terminate Anthropic’s $200 million contract and impose designations that could affect the company’s financial health.
Claude became the first AI model to operate on a classified military network. Last summer, Anthropic signed a deal worth up to $200 million. Other big companies in the AI space, such as OpenAI, contract with the Department of Defense only on unclassified networks.
Anthropic usage guidelines prohibit Claude’s use of mass surveillance and autonomous weapons.
“This conflict comes at a troubling time. On the one hand, DoD users love Anthropic, love Claude, and say its usage restrictions have never been enforced, at least from what I’ve heard.”
But the Pentagon doesn’t want to rely on private companies’ policies. “Tactical operations involving carve-outs cannot be conducted,” a spokesperson for the department told CNN. “The legality is the responsibility of the end user, the Department of Defense.”
From a departmental perspective, it is important to avoid situations where you need to ask for permission and remove “prior permission” for certain controls.
Parting ways with Anthropic could cause major headaches for the Pentagon. Internal systems using Claude will likely need to be replaced. At the same time, one official said Elon Musk’s Grok AI is “ready for use in sensitive environments” but is not considered a solution as advanced as Claude’s.
Losing the $200 million contract doesn’t eliminate Anthropic’s potential, which is valued at about $380 billion. But the bigger risk is the “supply chain risk” designation, which requires defense contractors to certify that they have no ties to Anthropic in doing business with the Department of Defense. Much of Anthropic’s success comes from contracts with large companies, some of which may have their own contracts with the Department of Defense.
“This means that Anthropic’s existing customer base could shrink or disappear due to government contracts or a desire to acquire customers in the future,” said Adam Connor, vice president of technology policy at the American Progress Center in Washington.
NVIDIA CEO Jensen Huang expressed hope that the Pentagon and Anthropic can reach an agreement, but added that it’s not the end of the world if the compromise fails, as there are other players in the market and Anthropic has other customers as well.
Earlier this week, the Pentagon considered applying the Defense Production Act of 1950 (DPA), a provision that gives the president broad powers. It is unclear whether the Department of Defense can designate anthropics as a “supply chain risk” and apply the DPA at the same time.
Humanity is not alone in its conflicts with governments. The Pentagon threat sends a signal to other AI companies. If you want to continue working with government agencies, don’t limit the use of AI technology.
“In a broader sense, I think this sends a signal to other AI companies not to try to put limits on their use of AI.”
If the Pentagon decides not to agree, it could terminate the contract and turn to another company with a similar AI model, said Alan Rosenstein, a law professor at the University of Minnesota.
“The government really wants to keep using Anthropic’s technology, and they’re just using every tool they can to influence it,” he concluded. “This is a very powerful means of influence.”
The Pentagon said it would terminate the contract if Anthropic refused by 5:01 p.m., and Anthropic cited “supply chain risks.” It is not yet clear whether a public notice will be issued, whether Claude AI will immediately disappear from military systems, or what will replace it. Questions also remain about how other defense contractors working with Anthropic will respond going forward. If a foreign company was deemed a risk, the company was given a period to sever ties.
If the Pentagon were to take radical steps, it would be an unprecedented escalation given one of the most famous and powerful advances in AI. “Taking away an in-house AI champion at a time when the White House is saying that the AI race with China is akin to the Soviet-era space race, you wouldn’t want to destroy one of the industry’s most valuable jewels by such an act,” the scenario reconstruction expert noted. “There are better ways to resolve this conflict than the absolutist path chosen by the regime.”
Chris Isidore contributed to this material.
