Google is negotiating to bring Gemini AI into classified Defense Department systems, testing how far AI contracts can reach in national security.
Google is negotiating an expansion of its existing Pentagon contract to allow Gemini to run inside classified networks, The Verge and Newsweek reported Thursday, citing Reuters. The talks mark a direct reversal: Google's current Defense Department agreement restricts Gemini to unclassified environments only.
A Pentagon spokesperson confirmed the direction of travel. "The department will continue to rapidly deploy frontier artificial intelligence capabilities to the warfighter through strong industry partnerships across all classification levels," the official told Newsweek. Google had not responded to requests for comment by publication time.
The proposed contract terms apparently mirror language OpenAI secured in its own Pentagon deal earlier this year, according to The Verge. That agreement covers OpenAI's models for "all lawful purposes," and despite provisions that seemed to bar fully autonomous lethal weapons and mass domestic surveillance, legal analysts warned the phrasing was broad enough to permit those applications anyway. Whether Google's version includes tighter limits is not yet public.
The classified push
Senior defense officials have framed artificial intelligence as a strategic necessity, repeatedly citing China's development pace as the rationale for speed over caution. That urgency has produced a multi-vendor strategy. Anthropic's Claude models have been running inside the military for data analysis and decision-support functions, explicitly not for autonomous combat roles, while the department simultaneously courts other suppliers to avoid over-dependence on any single company.
Exactly what strained the Pentagon's relationship with Anthropic is not detailed in the reporting. On the same day these talks became public, Anthropic separately released Claude Opus 4.7, its most capable publicly available model, per Mashable, suggesting the company remains commercially active regardless of any government friction. The defense AI market is not a winner-take-all contest.
Google's return to overt military AI work carries historical weight. After roughly 3,000 employees signed a protest letter in 2018, the company abandoned its Project Maven drone imaging contract and pledged not to build AI designed to cause harm. The following years brought quiet re-entry into defense contracting. Extending Gemini into classified settings would be the most explicit signal yet that the post-Maven era of internal restraint is over.
What it means for the market
Every major frontier AI lab is now converging on government access as a revenue pillar. OpenAI closed a $122 billion funding round at an $852 billion valuation, per the company's announcement, with defense contracts framed as part of its commercial expansion thesis. Google's Gemini push follows the same logic.
The "all lawful purposes" clause recurring in Pentagon AI contracts warrants scrutiny that has not yet materialized in Congress. The European Union's Artificial Intelligence Act established binding rules for high-risk applications in national security-adjacent contexts, but US federal procurement operates without an equivalent framework. Contract language remains the primary constraint on how these models are used in settings the public cannot inspect - a thin guardrail for systems operating at classification levels that preclude outside review.
Those guardrails may prove thinner still. The question facing Google, the Pentagon, and observers alike is whether an AI system trusted with classified data will eventually be authorized for decisions with no human in the loop at all.
FAQ
What is the Pentagon's current agreement with Google on AI?
Google holds a contract that lets the Defense Department use Gemini for "all lawful purposes," but only in unclassified settings. The new negotiations aim to extend that access into classified networks.
Why does the "all lawful purposes" language matter?
Legal analysts who reviewed the identical clause in OpenAI's Pentagon contract concluded it could permit uses - including autonomous lethal weapons and mass domestic surveillance - that the agreement appeared to exclude. The same ambiguity would apply to Google's deal if the wording matches.
What happened between Anthropic and the Pentagon?
Reporting indicates tensions led officials to seek alternatives, though neither The Verge nor Newsweek specified the source of the dispute. Anthropic continues to ship new frontier models and has not publicly addressed any rift with the Defense Department.
How does this compare to Google's Project Maven exit in 2018?
Project Maven was a narrowly scoped image-recognition contract for drone footage. The current Gemini talks involve broad deployment across classified infrastructure, making the scale of the potential policy reversal considerably larger.