Industry·3 min read
By BitsMindsSource: Fortune

1,000 Google Workers Sign Letter Against Pentagon AI Deal — But This Isn’t Project Maven

Roughly a thousand Google employees have signed an open letter opposing the company’s Pentagon Gemini deal. Unlike 2018’s Project Maven, leadership is digging in — and experts say staff leverage has materially eroded since then.

1,000 Google Workers Sign Letter Against Pentagon AI Deal — But This Isn’t Project Maven
Share:

An open letter from Google employees opposing the company’s decision to allow Gemini models inside U.S. military classified networks has gathered roughly 1,000 signatures, up from about 600 when the document first surfaced last week, according to a Fortune report published May 4. The dissent is the largest internal protest at Google over a defense contract since 2018’s Project Maven — but the parallels mostly highlight how much the balance of power inside the company has shifted.

The contract itself is unusually broad. Google has agreed to make Gemini available for “any lawful purpose” inside Pentagon classified networks, with a provision requiring the company to modify safety filters at government request. Researcher Alex Turner, who reviewed the public elements of the deal, summarized it as Google affirming it cannot veto military usage and committing to adjust safety guardrails on demand — language that goes well beyond the narrower carveouts in OpenAI’s and xAI’s comparable agreements. Legal expert Charlie Bullock and AI policy researcher Seán Ó héIgeartaigh assessed the Google contract as “strictly weaker” than OpenAI’s on the questions employees most worry about, particularly mass domestic surveillance and autonomous weapons use.

Where 2018 produced a corporate retreat, 2026 is producing the opposite. Google leadership issued a memo making clear the company “proudly” works with the U.S. military and intends to deepen the partnership, not exit it. Former Google engineer Laura Nolan, who resigned in protest of Project Maven, told Fortune that internal organizing has become structurally harder: “Staff in tech have also never been particularly well organized,” she said, noting that Google dismantled the internal mailing lists, all-hands forums, and cross-team chat channels that had given Maven critics their megaphone seven years ago.

The macroeconomic backdrop has shifted too. In 2018, AI talent had near-unlimited optionality, the labor market was tight, and a credible threat to walk out actually moved executive calculus. In 2026, AI labs are running aggressive efficiency programs, the AI infrastructure spending boom has tightened operating discipline across hyperscalers, and military and intelligence contracts have become a strategically core revenue line rather than a peripheral one. Google’s leadership appears to have concluded that the cost of losing Pentagon business now exceeds the cost of internal dissent — a reversal of the 2018 calculus.

Whether the protest gains further traction will be the test. Maven scaled from internal letter to public petition to executive backdown over the course of months, partly because key engineers were willing to resign and partly because Google’s public-facing AI ethics commitments at the time were fresh and politically salient. Neither condition fully holds today. The current letter’s signatories are betting they can force a similar climbdown without those tailwinds; the company’s response so far suggests it has bet they cannot.

Want AI news before everyone else?

The morning's most important AI stories, straight to your inbox. No fluff.

Related Articles

417–3 in the House. One objection in the Senate. Editorial illustration of the stalled Ratepayer Protection Act, H.R. 9340. A large 417–3 House vote meets a single red senator and a closed stop line. Below, AI data centres drawing 100 MW or more connect through a transmission tower to a home and an electricity bill. Amber conduits link the industrial load, grid upgrades and household costs. The proposed large-load standard would assign full incremental grid-upgrade costs to data centres, but a senator blocked unanimous consent, objecting that states were only asked to consider the standard. H.R. 9340 RATEPAYER PROTECTION ACT 417–3 HOUSE VOTE ONE SENATOR. BILL BLOCKED. The objection: states only asked to “consider”. AI DATA CENTRES 100 MW+ GRID UPGRADES WHO PAYS? BILL $ HOUSEHOLD BILLS BITSMINDS.COM
Industry

House Votes 417–3 to Make Data Centers Pay for the Grid

Hacktron: from an image upload to an internal repository An oversized HEIC photo file is connected by an electric green line to two glass security gates, labelled Forum and SSO, then to a dark repository cabinet bearing the OpenAI emblem and a pull-request symbol. A separate Claude Opus 5 plaque credits the model used by the researchers. The top caption says disclosed and patched. This is a conceptual illustration of Hacktron's reported security research and vulnerability chain. The two gates represent the two reported flaws, and their open padlocks are a visual metaphor. Original vector editorial illustration for BitsMinds, hacktron-claude-opus-5-openai-exploit-chain. 18 September 2026. Conceptual architecture based on the article, not a real interface or technical system diagram. HACKTRON / SECURITY RESEARCH DISCLOSED & PATCHED .HEIC IMAGE UPLOAD 01 FORUM IMAGE PROCESSING 02 SSO SIGN-IN TRUST INTERNAL REPO PULL REQUEST OPENED CLAUDE OPUS 5 RESEARCH ASSISTANT TWO FLAWS. ONE CHAIN. BITSMINDS.COM
Industry

Claude Opus 5 Wrote the Exploit That Broke Into OpenAI

Pew's global AI survey: uncertainty about the future of work An empty ochre office chair faces a computer in a quiet evening office. Beyond the window, buildings of unequal heights suggest an uneven economy. A large globe on the left is surrounded by 37 separate markers: 34 amber and three neutral. They represent surveyed publics, not individual respondents, countries mapped to locations, job counts or measured job losses. The caption states that in 34 of 37 publics, more people expect job losses than gains. This is an editorial illustration of public expectations reported in the Pew survey, not a prediction of what AI will actually do. Original vector illustration for BitsMinds, pew-global-ai-survey-2026-jobs-inequality. 17 September 2026. Survey finding, not observed labour-market outcomes. PEW / GLOBAL AI SURVEY 34 / 37 PUBLICS More expect job losses than gains AI BITSMINDS.COM
Industry

Most of the World Expects AI to Cost Jobs, Pew Finds