Google is in talks with the U.S. Department of Defense over an agreement that would allow the Pentagon to deploy its Gemini AI models in classified environments, according to reporting from The Information citing two people with direct knowledge of the discussions. Neither Alphabet nor the Pentagon responded to requests for comment. A Pentagon spokesperson said in a separate statement that the department “will continue to rapidly deploy frontier AI capabilities to the warfighter through strong industry partnerships across all classification levels.”

Under the proposed terms, the Pentagon would be able to use Gemini for “all lawful uses.” Google is separately pushing for contract language that would prohibit deploying its models for domestic mass surveillance or for autonomous weapons used without appropriate human control. Those restrictions closely track the terms OpenAI negotiated in its own Pentagon agreement earlier in 2026, and they are noticeably narrower than the blanket use-case restrictions Anthropic insisted on before the Pentagon formally designated it a supply-chain risk in March and ordered a six-month phase-out of Anthropic products across the department.

A reversal Google spent a decade building

The deal, if concluded, would mark the most significant break from Google’s post-Project Maven military stance since that controversy began. In 2018, thousands of Google employees protested the company’s involvement in Project Maven, the Defense Department drone-imagery program, citing Google’s then-operative “don’t be evil” motto. Google declined to renew the Maven contract, published a set of AI principles that explicitly excluded weapons and surveillance applications, and for much of the following decade ceded classified national-security work to Amazon, Microsoft, and Palantir.

Google began rewriting that posture in stages. In February, the company revised its AI principles to permit weapons and surveillance applications under appropriate oversight, a move that prompted an open letter from roughly 900 Google employees opposing the change. On March 10, Google announced that it would supply the Defense Department with eight pre-built Gemini-powered AI agents running on unclassified DoD networks, covering tasks such as meeting summarization, budget construction, and checking proposed actions against the national defense strategy. Google has said 1.2 million DoD personnel have accessed Gemini through the department’s GenAI.mil portal since December, generating 40 million prompts and uploading more than four million documents. Only 26,000 have completed formal AI training, though the company says future sessions are fully booked.

The proposed classified deal is the logical next step along that sequence, and it represents the first time Google’s models would operate inside the Pentagon’s classified environments. Amazon Web Services and Microsoft Azure have operated classified government clouds for years, and both have extensive experience clearing personnel, facilities, and software for work above the unclassified threshold. Google’s cloud has not historically carried the same accreditations at scale. Clearing Gemini into that environment would put Google in direct competition with OpenAI, Microsoft, and Amazon for classified AI workloads, not just unclassified productivity agents.

The pattern the Pentagon is establishing

The more interesting story is what this says about the Defense Department’s posture toward frontier AI labs as a group.

Six months ago, the three frontier labs most likely to power classified U.S. national-security AI work were arguably OpenAI, Anthropic, and Google, in roughly that order of depth of commercial relationship. Today the picture looks different. OpenAI signed a Pentagon deal early in the year and accepted contract terms that keep its restrictions narrow enough for the department to operate inside. Anthropic refused to loosen its safety posture, was publicly designated a supply-chain risk, and is now subject to a six-month phase-out across DoD, a decision it is challenging in court. Google rewrote its AI principles, began with unclassified agent work, and is now negotiating for classified access on OpenAI-style terms.

Taken together, the three outcomes describe a coherent procurement strategy. The OpenAI agreement effectively set the acceptable restriction template: narrow, specific carve-outs on domestic surveillance and autonomous-weapons edge cases, paired with a broad permission to use the model for “all lawful uses” otherwise. Labs that accept those terms become eligible for classified work. Labs that push for broader restrictions on how their models can be used by the U.S. government are pushed out, and the Anthropic blacklisting is the standing example of what that looks like in practice. The Google talks are the clearest sign yet that other frontier labs have read the incentive structure and adjusted.

The implications run in two directions.

On the infrastructure side, a classified Gemini deployment is a meaningful operational lift rather than a paper commitment. It implies Google investing in classified cloud enclaves, clearing personnel and facilities to the relevant levels, and building model-serving infrastructure that can operate without phoning home to standard Google Cloud control planes. That is expensive capacity, which is why it has historically concentrated at AWS, Azure, and a handful of specialized contractors. If Google clears that bar, the set of companies that can run frontier AI inside the U.S. classified fabric expands from roughly two to roughly three, which matters for resilience, competitive pricing, and the pace at which DoD can actually adopt what it buys.

On the governance and geopolitical side, the Pentagon now has a demonstrated template for moving frontier AI labs toward its preferred contract structure, and it has demonstrated the willingness to use blacklisting to enforce it. That is a very different environment from the one labs were operating in even a year ago, when AI safety policies were a matter of internal deliberation and public positioning. They are now a procurement variable. Labs that want access to classified U.S. government work will effectively be choosing between the OpenAI template and the Anthropic outcome, and their public principles will increasingly read as a signal of which choice they have already made.

Whether Google’s proposed restrictions on surveillance and autonomous weapons actually make it into the final contract is still unclear, and the talks could collapse. But the direction of travel is visible. Google spent most of a decade treating distance from the Pentagon as a core part of its corporate identity. It is now spending 2026 closing that distance, in public, in stages, and on terms the department has effectively set.

Sources: Newsweek, Sherwood News, SRN News (summarizing The Information), Engadget, Japan Times, Built In

–
By the Control Plane Editorial Team