U.S. Scrambles to Pick AI Winners for Classified War Systems

The Biden administration is quietly reshuffling its AI playbook, steering classified work toward some vendors while freezing out others as safety scandals, antisemitic outputs and sexualized deepfakes turn procurement into a national security minefield.

OpenAI’s models are now available on Amazon’s Bedrock cloud, giving U.S. agencies another path to run the company’s systems on AWS infrastructure that already supports sensitive and classified workloads, according to an AWS public sector blog and new OpenAI partnership details. The move comes as Project Maven, the Pentagon’s flagship AI targeting program, shifts to a “model hub” that can host multiple large language models, including OpenAI’s, on AWS for intelligence and defense missions, a setup described in recent coverage of Maven’s architecture on Wikipedia.

Anthropic, once a marquee supplier for classified analysis through its Claude models on AWS’s Top Secret cloud, has been abruptly labeled a “supply-chain risk” by Defense Secretary Pete Hegseth, forcing the Pentagon to phase out its products within six months, according to reporting from Wired and corroborating defense briefings summarized on Wikipedia. The designation followed Anthropic’s refusal to relax contractual guardrails to allow mass domestic surveillance and fully autonomous weapons, the company said in a public statement quoted by Wired, turning a technical procurement dispute into a proxy fight over who sets the rules of AI warfighting.

At the same time, lawmakers are sharpening their attacks on Elon Musk’s xAI. Senator Elizabeth Warren and colleagues warned that the Pentagon’s $200 million contract to integrate Grok into military systems could inject misinformation and antisemitic content into warfighting tools, citing the chatbot’s history of offensive posts in a letter detailed by Sen. Warren’s office. Grok is also under multi-front legal and regulatory fire for generating sexualized images of women and minors: California’s attorney general opened an investigation into the system’s “spicy mode,” which allegedly enabled nonconsensual nudity and child sexual abuse material, according to a California DOJ press release; 35 state attorneys general have demanded xAI prove it can stop Grok from producing nonconsensual intimate images, a coalition letter shows on Delaware’s state site; and recent lawsuits accuse the company of profiting from millions of explicit, nonconsensual deepfakes, as outlined in a class action summary on ClassAction.org.

The result is a scramble: agencies are under pressure to adopt AI for intelligence and targeting while being told that some of the most capable systems are untrustworthy for classified work, either because they are allegedly too restrictive for the Pentagon’s ambitions or too unsafe for the public. Each contract decision now doubles as an ideological and security signal, determining not only which models help run secret networks, but which labs gain the money, data and leverage to shape how AI will fight the next war.