Google signs classified AI deal with the Pentagon, joining OpenAI and xAI

TL;DR:

• Google has signed a contract letting the US Department of Defense — recently rebranded the “Department of War” by President Trump — use its AI models for “any lawful government purpose” on classified networks, according to The Information.
• The contract requires Google to help adjust safety settings and filters at the government’s request, while excluding domestic mass surveillance and autonomous weapons “without appropriate human oversight”.
• The deal places Google alongside OpenAI and Elon Musk’s xAI as classified AI suppliers; Anthropic was previously designated a Pentagon “supply-chain risk” after refusing to relax similar guardrails — a divergence with implications for how UK MoD and AISI think about analogous procurement.

Google has agreed a deal allowing the US Department of Defense to use its artificial-intelligence models on classified networks for “any lawful government purpose”, The Information reported on Tuesday. According to a person familiar with the agreement, the contract sits alongside existing Pentagon arrangements with OpenAI and Elon Musk’s xAI, formalising Google as a classified-tier AI supplier. Reuters has not independently verified the report. The Department of Defense — recently renamed the Department of War by President Donald Trump — did not respond to requests for comment.

Classified networks are used for sensitive work including mission planning and weapons targeting. The Pentagon last year signed agreements worth up to $200 million each with major AI labs, including Anthropic, OpenAI and Google. Reuters has previously reported that the Department was pushing top AI companies to make their tools available on classified networks without the standard restrictions they apply to commercial users.

What is in the contract

The Google deal includes language stating “the parties agree that the AI System is not intended for, and should not be used for, domestic mass surveillance or autonomous weapons (including target selection) without appropriate human oversight and control”. A separate clause, however, says the agreement does not give Google a right to control or veto lawful government operational decision-making. Google is also required to help the government adjust the company’s AI safety settings and filters on request — a meaningful concession that effectively makes Google’s safety configuration negotiable inside a classified environment.

A Google spokesperson said the company supports government agencies on both classified and unclassified projects and remains committed to the consensus that AI should not be used for domestic mass surveillance or autonomous weaponry without human oversight. The Pentagon has said publicly that it has no interest in mass surveillance of Americans or autonomous weapons but wants “any lawful use” of AI permitted.

Anthropic, and the UK contrast

The Reuters report makes Anthropic’s earlier breach with the Pentagon look more consequential. After refusing to remove guardrails against autonomous weapons or domestic surveillance, the Claude developer was designated a supply-chain risk by the Department — a label that complicates federal procurement and raises real commercial cost on safety stances. Google, OpenAI and xAI, by contrast, have all chosen to operate inside the new framework.

For UK readers, the contrast matters. The UK Ministry of Defence has been quietly building its own AI capability, and the AI Security Institute partners with Anthropic, Google DeepMind and OpenAI on red-teaming pre-release models. UK procurement teams now have to think harder about an emerging two-track AI vendor market: a US classified track where safety filters are tuneable by the customer, and a separate commercial-and-research track on which UK-sensitive workloads typically run. Defence-adjacent UK firms — particularly those running on Google Cloud or considering Vertex AI for sensitive workloads — should review whether their commercial terms are isolated from the Pentagon configuration.

Looking forward

The deeper question is whether tunable safety filters bleed back into commercial offerings. If governments routinely renegotiate guardrails in classified environments, civilian customers will eventually demand similar control. UK regulators and the AISI should treat the Google-Pentagon framework as a leading indicator for how Western governments will approach AI procurement, and consider what UK-specific procurement language is needed to keep autonomous-weapons and mass-surveillance restrictions intact when models cross the Atlantic.