Google has entered into a classified agreement with the Pentagon regarding AI.

      **TL;DR** Google entered into a confidential AI agreement with the Pentagon for “any lawful government purpose” just a day after over 580 employees petitioned CEO Sundar Pichai to decline such an arrangement. The contract includes certain advisory limits (prohibiting mass surveillance and autonomous weapons without human intervention) but allows the government to request changes to safety protocols. On the same day, it was reported that Google had surreptitiously withdrawn from a $100 million drone swarm competition in February following an ethics assessment. Google is attempting to distinguish between offering general AI access and developing specific weaponry, but this distinction may be negligible within classified networks.

      Google confirmed that it has signed an agreement permitting the Pentagon to utilize its Gemini AI models for secret military operations under conditions allowing for “any lawful government purpose,” a move that came one day after more than 580 employees urged Pichai to reject this type of partnership. This agreement allows the Department of Defense to access Google’s AI systems via API on classified networks, furthering an existing relationship that already enables Gemini access for three million Pentagon staff on unclassified systems. The contract stipulates that “the AI System is not intended for, and should not be used for, domestic mass surveillance or autonomous weapons (including target selection) without suitable human oversight and control.” Concurrently, Bloomberg reported that Google had quietly opted out of a $100 million challenge to develop voice-controlled drone swarms for the Pentagon, having withdrawn in February after conducting an internal ethics review despite being a frontrunner in the competition. Google officially attributed its withdrawal to a lack of resources. The company is delineating a boundary, albeit not one that aligns with employee expectations.

      **The deal**

      The classified AI agreement serves as an extension of Google’s existing Pentagon contract, providing API access instead of developing specialized military applications. A representative from Google’s Public Sector confirmed this arrangement. The Pentagon is able to connect to Google’s software on classified networks—isolated systems unconnected to the public internet that manage mission planning, intelligence analysis, and weapon targeting. The language permitting “any lawful government purpose” aligns Google with organizations like OpenAI and Elon Musk’s xAI, both of which have also established classified AI agreements with the Pentagon. The government may request modifications to Google’s AI safety settings and content filters, essentially granting the Pentagon the leeway to alter the safeguards that Google’s own researchers built into the models.

      The proclaimed limitations—prohibiting mass surveillance and autonomous weapons without human oversight—reflect similar stipulations negotiated in OpenAI’s Pentagon deal. However, the enforcement mechanism highlighted by Google’s employees as inadequate remains unchanged: on air-gapped classified networks, Google lacks visibility into the queries run, outputs generated, or decisions made based on those outputs. The phrase “should not be used for” is advisory rather than a contractual ban, and “appropriate human oversight and control” lacks a clear definition. The employees remarked that the sole way to ensure that Google avoids involvement in such harmful applications is to reject classified workloads altogether. Google opted to accept them with provisions that the employees previously argued were unenforceable. Pichai launched Cloud Next 2026 by highlighting 750 million Gemini users and a backlog of $240 billion. The same Gemini infrastructure serving these users is now being made accessible to classified military networks where its deployment cannot be monitored by anyone outside the Pentagon.

      **The withdrawal**

      The drone swarm situation adds another layer to this narrative. Google had made significant progress in a $100 million Pentagon prize competition aimed at developing technology for voice-controlled autonomous drone swarms, which would translate spoken commands like “left” into digital instructions for the drones. On February 11, 2026, the company informed the government of its decision to cease participation. Officially, Google cited resource constraints, although recorded information reviewed by Bloomberg indicated that the decision was influenced by an internal ethics review. This withdrawal is reminiscent of Project Maven in 2018 when around 4,000 Google employees petitioned against the AI analysis of drone video feeds, leading Google to allow that contract to expire, which Palantir subsequently took over. The Maven contract was valued at millions of dollars, while Palantir’s investment in Maven has since ballooned to $13 billion.

      The contrast is striking. On the same day that Google announced a deal granting the Pentagon classified access to Gemini for “any lawful government purpose,” it also revealed its withdrawal from a plan to utilize its AI for controlling autonomous drone swarms. Google is willing to deploy its most advanced AI models on classified networks without monitoring their application, but refrained from developing voice-controlled drone swarms. This distinction is significant within Google's internal ethical guidelines: providing API access to general AI models is seen as one step removed from weapon applications, despite the fact that these models will be used on systems associated with weapon targeting. Conversely, creating technology specifically aimed at controlling drone swarms is treated as a direct application for military use that did