Google signs Pentagon classified-AI deal as employee dissent grows

TL;DR: Google has signed an agreement letting the Pentagon use its AI models for classified work, joining OpenAI and Elon Musk’s xAI on equivalent contracts. More than 600 Google employees have signed an open letter opposing the deal. The Pentagon awarded contracts of up to $200 million each to leading labs in 2025. For UK defence and allied governments, the deal sets a precedent for what frontier-lab access to classified networks now looks like.

The agreement permits Pentagon use of Google AI for “any lawful government purpose”, The Information reported and The Guardian summarised. Classified networks are used for mission planning and weapons targeting, among other sensitive workflows. The contract reportedly requires Google to help adjust the model’s safety settings and filters at government request.

Where the lines are drawn — and where they are not

Contractual language explicitly excludes domestic mass surveillance and autonomous weapons “without appropriate human oversight and control”. But the deal also denies Google a veto over lawful operational decisions — meaning the government, not the vendor, sits in the chair on use-case interpretation. This is the same architecture Anthropic refused earlier in the year: the company kept guardrails on autonomous weapons and domestic surveillance, and the Pentagon designated it a supply-chain risk in response.

Google’s reversal is striking. Alphabet lifted its ban on AI for weapons and surveillance work last year, removing language that promised the firm would not pursue technologies likely to cause overall harm. The 2018 employee revolt over Project Maven — which led Google to drop the contract and Palantir to take it over — looks, from this distance, like a one-off rather than a settled stance.

Looking forward

For UK MoD and allied procurement teams, the practical implication is that frontier-lab access to classified workloads is now a market expectation, not a concession. The UK’s defence-AI buyers can negotiate from a position where access has been priced and contractual terms have been published. The harder question is governance: the Anthropic carve-out shows vendors will still accept commercial cost to protect specific guardrails, which means UK procurement should specify which lab posture matters before signing. The 600 Google signatories may not move the contract, but they document the talent-side cost of choosing the more permissive vendor.