Google’s Pentagon AI deal opens classified use with looser veto power

Google has moved one step deeper into military AI, and the real story is not just access. It is control. A Reuters report published April 28 says Google signed a Pentagon deal that allows its models to be used for classified work and, crucially, lets the government request adjustments to safety settings and filters.

That matters because frontier AI labs have spent months drawing public red lines around domestic surveillance and autonomous weapons. According to Reuters, citing The Information, the Google agreement says the system is not intended for domestic mass surveillance or autonomous weapons without appropriate human oversight. But the same report says the deal does not give Google the right to control or veto lawful government operational decisions. In plain terms, the company can state principles, yet once its model is inside classified environments, operational leverage may shift toward the customer.

The Pentagon uses classified networks for highly sensitive work, including mission planning and weapons targeting. Reuters adds that the Defense Department signed AI agreements worth up to $200 million each in 2025 with Anthropic, OpenAI, and Google, and has pushed major labs to make models available on classified systems with fewer restrictions than in consumer or enterprise products. Google told Reuters it remains committed to the consensus against domestic mass surveillance and autonomous weaponry without human oversight.

The timing sharpens the stakes. Anthropic has already clashed with the Pentagon after refusing to remove guardrails tied to autonomous weapons and domestic surveillance. Google’s path looks different: accept the defense demand signal, keep the principles language, and rely on contractual and operational process rather than a hard veto. That is a major shift in how commercial AI governance works once government procurement enters the picture.

For the wider AI market, this is bigger than one defense deal. Classified deployment forces labs to answer a harder question than whether they should work with the military. The harder question is how much control survives after the model is deployed. Reuters’ report suggests that answer may be narrower than the public language many labs use. The source article is here.