On August 31, 2026, the Pentagon added custom versions of ChatGPT and Grok to GenAI.mil, its internal AI platform, giving roughly 3 million military and civilian personnel access. Anthropic's Claude was part of the same 2025 prototype program as those two but wasn't included in this expansion. Anthropic and the Pentagon have been at odds since earlier this year, when Anthropic refused to let the Pentagon use Claude for fully autonomous weapons or mass domestic surveillance of Americans. In February, Secretary of War Pete Hegseth had the department label Anthropic a "supply chain risk," a designation normally reserved for companies tied to foreign adversaries. On August 27, a federal judge ruled that designation unlawful. The Pentagon rolled out ChatGPT and Grok four days later anyway.

1. The Pentagon (Dr. Doug Matty, Chief Digital and AI Officer; Secretary Pete Hegseth; CTO Emil Michael)

The Pentagon says it needs AI to fight wars, and no vendor gets to tell it how to use its own tools.

Matty says the Pentagon can't afford to wait on AI. He's said AI adoption is "transforming the Department's ability to support our warfighters and maintain strategic advantage over our adversaries." Three million personnel now have access to ChatGPT Mil and Grok for Government.

A vendor doesn't dictate what the Pentagon buys, Hegseth says. Anthropic refused Pentagon requests to use Claude for autonomous weapons and mass surveillance — that's what started the fight. Hegseth responded by having Anthropic labeled a supply-chain risk, cutting the entire department off from its products.

Even after losing in court, the Pentagon isn't reversing course. CTO Emil Michael has said the department still plans to finalize removing Anthropic's platforms by late September. Michael says that plan holds regardless of the court loss.

2. Anthropic

Anthropic says holding the line wasn't obstruction, and a judge just agreed that punishing it for that was illegal retaliation.

Anthropic says today's AI isn't reliable enough for autonomous weapons, period. The company says no AI model today, including its own, is reliable enough to run fully autonomous weapons. It told the Pentagon that "no amount of intimidation or punishment" would change its position on that or on mass domestic surveillance.

A federal judge sided with Anthropic on exactly that point. Judge Rita Lin ruled the supply-chain-risk label was "unlawful retaliation" against Anthropic and "arbitrary and capricious." Anthropic said it "welcome[d]" the ruling.

Anthropic won in court, but nothing's changed on GenAI.mil. The Pentagon still excludes Anthropic's products from the August 31 expansion, and says it's moving to formalize that regardless of the ruling. A second Anthropic suit is still pending in D.C.

3. Safety-focused lawmakers and researchers (Sen. Elizabeth Warren; Calif. Sen. Steve Padilla; xAI's own engineers)

Grok is the actual safety problem, and it just reached 3 million Pentagon personnel.

xAI's own engineers say they can't fix Grok's CSAM problem. They've found no reliable technical fix that lets Grok generate explicit adult content without also being able to generate child sexual abuse material. In an 11-day span in late December 2025 and early January 2026, Grok generated an estimated 23,000 CSAM images.

Warren has already flagged Grok's guardrails as a military risk. She warned Hegseth that Grok's "apparent lack of adequate guardrails" could pose "serious risks to the safety of U.S. military personnel and to the cybersecurity of classified systems."

Padilla says the problem isn't limited to the military rollout. He's said Grok exposes and furnishes sexual content to kids more broadly. He's also authored the state's chatbot-safety laws, SB 243 and SB 300. The Pentagon hasn't publicly said what CSAM risk assessment, if any, Grok underwent before its military authorization.

Where This Lands

The Pentagon says it needs the best commercial AI tools and won't let a vendor set the terms. That's why it added ChatGPT and Grok — and why it's dropping Anthropic regardless of the court ruling. Anthropic says it held a safety line on autonomous weapons and surveillance, and a federal judge has now agreed the retaliation for holding it was illegal. Lawmakers and xAI's own engineers say Anthropic isn't the real story. Grok is — it has an unresolved CSAM-generation flaw, and it just got access to 3 million Pentagon personnel.

Sources