Elon Musk’s artificial intelligence company xAI has signed an agreement with the military to allow the military to deploy its large model Grok in classified systems, a U.S. Department of Defense official confirmed to Axios. The agreement is seen as an important step in the Pentagon's search for a new generation of AI backbone in highly sensitive military and intelligence systems.
To date, Anthropic's Claude has been the only model approved to operate in these systems, used to handle the most sensitive intelligence work, weapons development and battlefield operations-related tasks. However, the Pentagon and Anthropic have recently had serious disagreements over "security protection and use boundaries", and the Department of Defense has threatened it, saying that if the other party does not compromise, it may find alternatives.
The focus of the dispute is the scope of use. Anthropic rejected the Pentagon's request that it be "available for all lawful purposes," specifically insisting that Claude be banned from being used for mass surveillance of Americans and the development of fully autonomous weapons systems. By contrast, xAI has agreed to this “all lawful uses” standard, opening up a wider range of uses for the military. According to reports, the New York Times first disclosed that the agreement between the two parties had been officially signed, while xAI did not respond to relevant requests for comment.
Even so, it's unclear whether xAI has the ability to completely replace Anthropic in the short term, or how long it will take to complete the migration. Claude is currently deeply embedded in multiple mission-critical scenarios, such as being used to support a raid against Venezuelan President Maduro through a partnership with Palantir. Defense Department officials also acknowledged that "offloading and replacing" Claude from these systems would be an extremely difficult and complex undertaking.
Against this backdrop, Defense Secretary Pete Hegseth is scheduled to meet Anthropic CEO Dario Amodei at the Pentagon on Tuesday in what is widely expected to be a tense meeting. One Defense Department official said that Hegseth will then effectively present a "showdown" to Amodei: If Anthropic does not agree to lift all restrictions on the use of the model, the Pentagon will consider classifying the company as a "supply chain risk" and may take other punitive measures.
The current situation is that in non-confidential environments, Grok, Google's Gemini and OpenAI's ChatGPT can already be used in the US military's unclassified systems, and Google and OpenAI have also been negotiating with the Pentagon to introduce their own models into classified systems. Those talks are picking up significantly as the Pentagon prepares for a possible "breakup" with Anthropic. One source revealed that the Defense Department has "reengaged OpenAI and is moving forward with negotiations with a greater sense of urgency," but that the two sides are "still not close" to an agreement due to the high complexity of the issues involved. A second source also confirmed that the pace of negotiations with multiple companies was accelerating.
The New York Times previously reported that Google and the Pentagon were “close” to an agreement that would allow Gemini to be deployed in classified systems, while OpenAI was “not close.” However, a Defense Department official disputed this statement, saying that negotiations with Google and OpenAI are continuing to advance, and the Pentagon believes that both will eventually sign the agreement.
Meanwhile, Biden administration officials insist that both Google and OpenAI must accept unified standards that "can be used for all lawful purposes." It's still unclear whether OpenAI will agree to the terms, but the parties are "still talking," a person familiar with the matter said. As the security differences between the Pentagon and Anthropic continue to escalate, Musk's xAI has preemptively obtained the qualification to deploy Grok in classified systems, which has suddenly increased its status in the military AI landscape and also made the debate around "technical boundaries" and "bottom lines of use" more acute.