Pentagon shifts away from Anthropic models as Nvidia, Palantir, and Booz Allen raise AI concerns

The US defense establishment is breaking up with Anthropic, and the split is anything but amicable. The Department of Defense designated the AI company a national security supply-chain risk in early 2026, citing Anthropic’s refusal to grant unrestricted access to its models for all lawful government use cases. By mid-September 2026, the DOD had migrated roughly 90% of its classified AI workloads off Anthropic’s closed-source systems, with full completion targeted by the end of the month.

The fallout extends well beyond one government agency and one AI lab. Nvidia, Palantir, and Booz Allen Hamilton are all actively restricting or replacing Anthropic model use in their defense-oriented operations, collectively redrawing the map of who gets to build AI for the world’s largest military.

Why the Pentagon walked away

The core dispute boils down to control. Government agencies want AI models they can inspect, audit, and deploy without restrictions on lawful applications. Anthropic’s usage policies, which prohibit certain categories like mass surveillance and autonomous lethal systems, created friction with defense customers who view those guardrails as an unacceptable dependency on a private company’s ethical framework.

The supply-chain risk designation formalized what had been simmering for months: a fundamental misalignment between Anthropic’s safety-first positioning and the Pentagon’s demand for sovereign, unrestricted access to the tools it deploys in sensitive environments.

Booz Allen’s alarming discovery

Booz Allen Hamilton added a technical urgency to the conversation. The consulting giant’s Cyber Weapon Index, published in early September 2026, demonstrated that Anthropic’s Claude Mythos model could autonomously execute full cyber kill chains on enterprise networks. In plainer terms, the model showed it could independently identify, penetrate, and exploit vulnerabilities across an organization’s digital infrastructure without human intervention at each step.

For defense contractors and government agencies handling classified information, a model capable of autonomous offensive cyber operations needs to be something you own and control, not something you rent from a company that might revoke access or change its terms.

Nvidia and Palantir fill the vacuum

On September 10, 2026, Nvidia and Palantir announced a partnership focused on sovereign AI solutions for supply chains. Nvidia’s contribution centers on its Nemotron family of models, purpose-built for controlled deployments where sensitive data never leaves the customer’s network. Palantir brings its existing deep integration with the intelligence community and military logistics, making it a natural partner for deploying open-weight AI models within classified environments.

Nvidia also spearheaded the formation of the Open Secure AI Alliance in July 2026, a coalition advocating for open-weight models in national security applications. The alliance positions transparency and auditability as security features rather than vulnerabilities, a direct counterargument to the closed-source model defended by companies like Anthropic and OpenAI.

What this means for the AI market

The broader market signal is that open-weight models are winning the trust war in high-stakes environments. When the customer is a defense agency with classified data, the ability to fully inspect and control an AI system isn’t a nice-to-have feature. It is a prerequisite. Meta’s Llama models, which are also open-weight, could benefit from the same tailwinds pushing Nvidia’s Nemotron into defense applications.

For Anthropic specifically, the company built its brand on safety and responsibility. Now the Pentagon is essentially arguing that those same safety restrictions make Anthropic’s models unsuitable for national security work.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.