Pentagon Signs 8 AI Companies for Classified Work While Shutting Out Anthropic

On May 1, 2026, the Department of Defense announced formal agreements with eight AI companies to deploy their frontier AI capabilities on the Pentagon's classified networks for "lawful operational use." According to TechCrunch, the companies — OpenAI, Google, SpaceX, Nvidia, Microsoft, Amazon Web Services, Oracle, and startup Reflection AI — will provide resources to deploy AI on IL6 (classified) and IL7 (top secret) environments.

Conspicuously absent: Anthropic, the company that was previously the Pentagon's primary classified AI partner and the first to deploy AI on classified networks. According to DefenseScoop, Anthropic remains designated a "supply chain risk" — a label typically reserved for foreign adversaries — after refusing to accept contract language that would allow "any lawful use" of its AI by the military.

The core dispute centers on contract language. The Pentagon wanted Anthropic to agree to "any lawful use" or "any lawful purpose" provisions. Anthropic refused to loosen its safety "red lines" around mass domestic surveillance and fully autonomous weapons systems, according to The Verge.

After talks collapsed, the Trump administration designated Anthropic a supply chain risk — a move Military Times notes was "the first such designation against an American firm." Defense Secretary Pete Hegseth moved "within days" to apply the label, per the BBC, and posted on X that "effective immediately, no contractor, supplier, or partner" doing business with the Pentagon may engage in any commercial activity with Anthropic.

Anthropic has filed two lawsuits challenging the designation and won a temporary injunction in March 2026. The legal battle is expected to go to court in September 2026.

The DOD press release stated the agreements will "streamline data synthesis, elevate situational understanding, and augment warfighter decision‑making" across three core tenets: "warfighting, intelligence, and enterprise operations," as quoted by TechCrunch.

According to DefenseScoop, the DOD's GenAI.mil platform — launched in December 2025 — has already seen over 1.3 million DOD personnel use it, generating "tens of millions of prompts and deploying hundreds of thousands of agents in only five months." GenAI.mil currently operates at IL5 (controlled unclassified); the new deals push into the IL6 and IL7 classified territory for the first time across multiple vendors.

Here's the twist: even as Anthropic is officially blacklisted, the NSA is already testing Anthropic's Mythos model to identify cybersecurity vulnerabilities in widely used software, including Microsoft products, according to Gizmodo. Mythos is Anthropic's most advanced AI model, released only to select organizations.

DOD CTO and Undersecretary of Defense for Research and Engineering Emil Michael told CNBC that Mythos represents a "separate national security moment," adding: "We have to make sure that our networks are hardened up, because that model has capabilities that are particular to finding cyber vulnerabilities and patching them," as reported by The Verge. Michael confirmed Anthropic is still a "supply chain risk" even as its most powerful model is being tested by the NSA — a paradox that highlights the tension between security concerns and the practical value of Anthropic's technology.

The Pentagon's deals reveal a sharp divide in how AI companies approach government work. Every company that signed accepted the "lawful use" framework — but with varying degrees of safety language. Gizmodo reports that OpenAI claims it maintained control over its safety stack and prohibited use of its AI for mass domestic surveillance or directing lethal autonomous weapons. But Google's deal explicitly states the company "does not confer any right to control or veto lawful Government operational decision‑making," per The Information, cited by Gizmodo.

At a Senate Armed Services Committee hearing on April 30, Defense Secretary Hegseth called Anthropic's leadership an "ideological lunatic who shouldn't have sole decision‑making over what we do." When asked whether there will "always be a human in the loop," Hegseth said "We follow the law and humans make decisions" and that AI is not currently "making lethal decisions," per Gizmodo.

President Trump told CNBC his administration had "some very good talks" with Anthropic and a future agreement was "possible," adding: "They're very smart, and I think they can be of great use."

Last summer, the Pentagon unveiled individual contracts with OpenAI, Anthropic, Google, and xAI — each worth up to $200 million — for "frontier AI projects," according to DefenseScoop. Anthropic had a $200 million deal to handle classified materials for the Pentagon, per The Verge. The financial terms of the new contracts were not disclosed.

DOD CTO and Undersecretary of Defense for Research and Engineering Emil Michael framed the expansion as a diversification strategy: "It's irresponsible to be reliant on any one partner. And we learned that that one partner didn't really want to work with us in the way we wanted to work with them," he told CNBC, as reported by DefenseScoop.

For developers and builders working with AI platforms, the Pentagon deals have several practical implications:

• Multi‑model classified infrastructure is here. The DOD explicitly framed these deals as preventing vendor lock‑in. Builders shipping products that integrate with government systems now need to support multiple AI backends, not just one.
• The "any lawful use" standard is becoming the norm for government AI procurement. Every major AI company except Anthropic has now accepted it. This sets a precedent that will shape future contracts at federal, state, and local levels.
• Anthropic's tools are still in use across government. The BBC notes Claude is still currently in use in many US government and defense agencies, creating a complex transitional period for agencies that built workflows around Anthropic's tools.
• OpenAI's safety claims are untested. While OpenAI says it maintains safety guardrails, Google's deal language explicitly surrenders veto power. Builders should watch how these safety commitments hold up under actual operational pressure.
• Automation bias is the real risk. Lauren Kahn, a former DOD policy advisor now at CSET, called the expansion "long overdue" but warned the next step must involve training operators against "automation bias" — the tendency to overdelegate to machines — per DefenseScoop.