Google Signs Classified Pentagon AI Deal as 950 Employees Revolt

Google has granted the U.S. Department of Defense access to its Gemini AI models for classified networks under terms that permit "any lawful government purpose," the company confirmed Tuesday. The contract is the third major AI deal the Pentagon has signed in the wake of Anthropic's refusal to allow similar use cases — and it lands in the middle of a public revolt by Google's own engineering staff.

A letter signed by 950 Google employees, addressed to CEO Sundar Pichai and circulated this week, asks the company to follow Anthropic's lead and bar classified military deployments without explicit guardrails against mass surveillance and lethal autonomous weapons.

What the Contract Covers

Under the new terms, the Department of Defense can use Google's Gemini frontier models on classified networks for an essentially unrestricted set of government purposes, subject only to existing U.S. law. That includes intelligence analysis, operational planning, and any other lawful application the DoD chooses to pursue.

The deal expands a relationship Google had previously kept narrow. Google Cloud already serves DoD workloads, but the older agreement carved out specific use cases and required additional review for classified deployments. The new contract removes those gates.

Pentagon AI chief Radha Plumb confirmed the work publicly Tuesday, saying the department's accelerated AI agenda required vendors who could move at the speed of operational need rather than litigate every classified mission. The DoD has been racing to bring frontier-model capabilities into intelligence and command-and-control systems, and the agency has made it clear it does not want vendor-imposed limits on how those systems are used.

Why Anthropic Said No

The contrast with Anthropic is now central to the story. When the Pentagon approached Anthropic on similar terms, the company refused, citing its responsible-scaling policy. Anthropic was willing to support defense and intelligence applications generally, but drew explicit lines against domestic mass surveillance and lethal autonomous weapons.

The DoD responded by branding Anthropic a "supply-chain risk" — a designation typically reserved for foreign adversary suppliers. Anthropic sued, and last month a federal judge granted an injunction blocking the designation while the case proceeds.

OpenAI and xAI both moved quickly to fill the gap, signing classified-use deals on terms close to what the Pentagon originally requested. Google is now the third major U.S. lab to do the same. That positions Anthropic as the only frontier AI company holding a public line against unrestricted defense use of its models.

The Employee Letter

The internal letter at Google was first reported by The Washington Post and confirmed by CBS News. It is direct in its language. Employees write that they "feel that our proximity to this technology creates a responsibility to highlight and prevent its most unethical and dangerous uses" and ask Pichai "to refuse to make our AI systems available for classified workloads."

Signers warn that "any lawful government purpose" framing leaves no real boundary. They cite mass surveillance, autonomous targeting, and operations targeting U.S. residents as the specific risks the contract fails to address. The letter explicitly asks Google to adopt a stance "in line with Anthropic's" — a notable inversion of the usual employee-to-management asks at large tech companies, which tend to invoke external standards rather than competitor policies.

The 950 signatures represent a fraction of Google's roughly 180,000 employees but draw heavily from the company's most senior AI research and infrastructure teams. Several signers were on the original 2018 Project Maven petition, the internal protest that ultimately led Google to drop a Pentagon drone-imagery contract.

What's Different This Time

The 2018 Maven episode ended with Google walking away from the contract and publishing AI principles that explicitly excluded weapons applications. Eight years later, the company's stance has shifted. The principles still exist on paper but have been amended multiple times, and Google now positions defense as a normal customer segment for Google Cloud and Gemini.

Several factors changed the calculus. The geopolitical environment is more hawkish, with explicit U.S. policy treating AI primacy as a national-security priority. Frontier-model revenue at scale increasingly depends on government and defense contracts. And the competitive pressure from OpenAI and xAI — both of which have built defense practices without internal blowback — has reshaped what executives see as table stakes.

Industry Implications

The DoD now has classified-use access to three of the four leading frontier model providers, with Anthropic the only holdout. For competitive procurement, that's a meaningful change. Programs that previously had to choose between a single vendor and a custom-built model can now run multi-vendor evaluations.

For Anthropic, the position is more nuanced than it might appear. The company has been winning enterprise and allied-government work that values its safety posture, including contracts with several U.S. agencies on terms that match its policies. Whether the DoD's "supply-chain risk" stance survives the lawsuit will determine how much of the broader federal market remains accessible.

For Google, the immediate question is internal. Maven cost the company senior researchers and reputational capital among the AI safety community. The new contract is larger and the stakes higher. Whether Pichai can hold the line on this deal without a wave of departures is the open question.

What's Next

Google has not commented publicly on the employee letter. The company's standard response to the Pentagon news has emphasized that all uses remain bounded by U.S. law and that classified workloads will run on Google-controlled infrastructure with appropriate oversight.

The Pentagon, for its part, has signaled that more contracts are coming. Plumb said this week that the DoD intends to integrate frontier models across multiple programs over the next 18 months, with a particular focus on intelligence analysis and decision support for combatant commands.

Bottom Line

Anthropic's refusal to grant the DoD unrestricted access has now produced three competitive contracts for its rivals — and an open revolt inside Google. The structural question for the AI industry is whether safety policies can survive in a market where the largest customer is willing to go elsewhere. For now, the answer at Google appears to be no, and 950 of its engineers are publicly saying that's the wrong call.