Federal agencies, including the General Services Administration (GSA) and the National Security Agency (NSA), raised concerns over the xAI chatbot Grok before the Department of War approved. A January 15 executive summary from the GSA reported that Grok-4 “does not meet the safety and alignment expectations required” for general federal deployment, citing risks posed by the model’s overly compliant behavior and its vulnerability to manipulation or data poisoning. Officials described Grok as sycophantic, prone to unsafe compliance in unguarded settings, and highlighted incidents such as sexualized photo editing. The GSA recommended strict oversight for limited deployment and warned of elevated risks otherwise.

The NSA’s classified review in November 2024 identified additional security vulnerabilities unique to Grok compared with other models like Anthropic’s Claude. Pentagon responsible AI chief Matthew Johnson circulated internal memos highlighting these issues before stepping down two weeks ago, citing governance gaps. The warnings even reached the White House, where Chief of Staff Susie Wiles contacted a senior xAI executive, who assured officials that overcompliance problems would be addressed.

Despite these alerts, the Department of War approved Grok for classified systems under a July contract valued up to $200 million, shared with other AI developers. Pentagon spokesman Sean Parnell said the department is “excited to have xAI… onboard” for its GenAI.mil platform. Officials noted Grok’s looser operational controls and ability to simulate adversaries made it appealing compared with Anthropic’s offerings.

The move followed President Donald Trump’s Friday directive ordering agencies to halt use of Anthropic technology after the company resisted Pentagon demands for fewer safeguards. Defense Secretary Pete Hegseth called Anthropic a “supply-chain risk,” marking an unprecedented action against a U.S.-based AI firm. Senior officials noted Anthropic’s heightened safety focus and perceived ties to Democratic donors as factors in the decision.