|
The Pentagon used xAI’s Grok to help bomb Iran. According to a court declaration from Cameron Stanley, the Defense Department’s chief digital and AI officer, a Grok model running on Palantir’s Maven Smart Systems supported the spring campaign against Iran, as part of a system that “enabled U.S. forces to deploy over 2,000 munitions to 2,000 distinct targets within 96 hours.” That is the disclosure, and it is a large one, and the government made it almost in passing.
It came out in an environmental lawsuit. The NAACP, with the Southern Environmental Law Center and Earthjustice, sued xAI over the gas turbines that power its Colossus data centers outside Memphis, near Boxtown, a historically Black neighborhood. The turbines, dozens of them, run without air permits, which xAI argues it does not need because they sit on trailers. The suit alleges they have degraded local air quality and asks the court to shut them down. The Justice Department intervened on xAI’s side, and its argument is that the turbines cannot be shut down because the AI they run helped bomb Iran.
Stated plainly, the government’s position is that an unpermitted power plant is national-security infrastructure, because one of the four AI models it runs for classified work is Grok, and Grok was in the targeting loop over Iran. Shutting the turbines, the DoJ wrote, would undermine “American national, economic, and energy security.” The environmental claims, whatever their merits, now have to be argued against the position that the facility is part of how the United States fights its wars.
This is also a lot to learn about Grok, which most people meet as the chatbot that posts on X. The same model is now documented running a wartime kill chain, which is a wide range of responsibilities for one product. And Grok belongs to xAI, now part of publicly traded SpaceX, so this is a disclosure about a public company: its chatbot directed live munitions, and the federal government treats its data centers as too important to regulate. That is, in its way, a competitive moat.
The contrast with last week is hard to miss. The same government switched off Anthropic’s most capable model three days after launch, on a national-security export order, because it might help someone find software flaws. Grok was used to destroy 2,000 targets and has been assigned a Justice Department legal team. One lab said its models were dangerous and asked to be regulated, and was regulated. The other built a model that was used in a war, and got a lawyer. Whatever the government is sorting frontier AI by, it is not how much harm a model can do, because on that measure these two are not close.
The lawsuit will be decided on its own terms. The larger fact is the one the filing disclosed almost in passing: a commercial AI model, the same Grok that posts on X, helped the United States put 2,000 munitions on 2,000 targets in 96 hours. A frontier chatbot has been used to run a bombing campaign, and the public found out from a court filing about gas turbines.
|
