Washington is quietly watching a high-stakes tech courtship unfold, as Google negotiates with the Pentagon over a deal that would let the Defense Department run the company’s Gemini AI models inside classified systems, according to people briefed on the talks. At the center of the discussions is proposed contract language that would limit how the models can be used, putting Google’s public AI safety promises under direct national security pressure.
What Google Is Negotiating
Google has proposed language that would prohibit its models from being used for domestic mass surveillance and from controlling autonomous weapons without “appropriate human control,” according to The Information. Company negotiators are also weighing an arrangement that would let the Pentagon deploy Gemini models inside classified environments, extending their reach far beyond current unclassified uses.
How Reuters Framed It
Reuters reported that Alphabet and the Pentagon are in talks over a classified AI agreement, following earlier reporting by The Information. According to Reuters, officials are exploring contract language that tries to square the Pentagon’s demand for operational flexibility with vendors’ public safety commitments.
Context: GenAI.mil And The Anthropic Fight
In March, the Pentagon launched GenAI.mil and made Google’s Gemini the first model available on its unclassified networks. Emil Michael said the department was “starting with unclassified” and would move to classified systems next, according to Bloomberg. That rollout, paired with a public standoff between the Pentagon and Anthropic over usage guardrails, has pushed procurement mechanics and safety limits to the center of every vendor conversation.
Why The Contract Language Matters
The core fight is over how far the Pentagon’s authority should reach inside commercial AI tools. Defense officials have pushed for broad “all lawful uses” wording that they argue is necessary to move quickly during crises. Companies, in turn, are trying to lock in red lines that would block domestic mass surveillance and fully autonomous lethal systems. Anthropic resisted that push and later sued after being labeled a "supply-chain risk," according to the Associated Press.
Legal And Procurement Risks
Legal experts say this tug-of-war could reshape how the federal government writes AI contracts for years to come. The Pentagon’s use of the rare "supply-chain risk" label against Anthropic has already triggered litigation and rattled procurement norms. The Information notes that if Google and the Pentagon end up preserving some company safeguards directly in classified contract language, it could set a powerful template for how commercial AI models operate inside secure government networks.
What’s Next
For now, talks are ongoing, and people familiar with the negotiations say there is still no agreement on which guardrails, if any, will survive inside classified contracts. Alphabet and the Pentagon did not immediately respond to requests for comment, according to Reuters.
