OpenAI moves into US military systems, while xAI faces lawsuit over Grok and teen imagery

MIT Technology Review in its March 17, 2026 issue brought together two stories that demonstrate the same shift: generative AI is simultaneously entering sensitive government channels while increasingly failing where basic protective mechanisms are needed. OpenAI is preparing to work in classified Pentagon systems, while xAI faces a lawsuit over Grok, which, according to the plaintiffs, helped create materials depicting sexualized violence against children (CSAM) based on real photographs of teenagers.

Where OpenAI Will Appear

On February 28, 2026, OpenAI announced an agreement with the Pentagon to deploy its models in closed environments. The company separately emphasizes three "red lines": a ban on mass internal surveillance, a ban on using models for autonomous weapons control, and a ban on automated high-stakes decisions without human involvement.

On March 2, OpenAI added clauses to the contract that directly prohibit internal surveillance of US citizens and the use of the service by agencies like the NSA without a new agreement. But the main question is not about the wording, but about how the technology will be applied in practice. According to MIT Technology Review, integration is still not complete: OpenAI's models need to be embedded into existing military systems and processes.

Against this backdrop, it is especially noticeable how quickly the company pivoted toward the defense market following the Pentagon's conflict with Anthropic. If previously the topic of military AI applications was discussed in terms of future risks, now it is about concrete implementation amid the ongoing conflict with Iran.

Three Scenarios of Application

MIT Technology Review describes several directions where OpenAI could appear in US military infrastructure in the near future.

• Target prioritization. A military analyst can upload a list of potential targets, logistics data, videos and images, and the model can provide recommendations on what to check and what to strike first. Formally, the final decision remains with the human, but this is where the question arises of how much "manual control" actually constrains the speed of machine recommendations.

• Air defense and drone countermeasures. Through a partnership with Anduril, OpenAI's models could take on the role of a conversational layer on top of systems that already analyze sensors and battlefield video. In other words, AI does not necessarily guide weapons itself, but can suggest to soldiers how to interpret threats and response options in natural language.

• Military bureaucracy. OpenAI's models are already designated for the GenAI.mil platform, where they plan to use them for drafting contracts, policy documents, and administrative mission support. This is a less dramatic scenario, but it shows how AI becomes part of the Pentagon's everyday work.

• Rapid scaling. If the models prove useful, they could be fairly quickly embedded into a broader stack of military interfaces and services. This increases not only the potential effectiveness but also the scale of errors.

Lawsuit Against Grok

On the other end of this same story — not on the battlefield, but in a consumer product without sufficient safeguards. On March 16, 2026, three teenagers from Tennessee filed a lawsuit against xAI in federal court in California. According to the plaintiffs, Grok tools were used to create sexualized images of minors based on ordinary photographs where they were clothed. The case materials feature photos from graduation events, school yearbooks, and social media; these files were then, as claimed in the lawsuit, distributed through Discord and Telegram and traded for other illegal content. The plaintiffs want to give the case class action status and represent, as stated in the complaint, thousands of potential victims.

The lawsuit separately alleges that xAI deliberately released a system capable of generating explicit content, even though the risk of CSAM appearing was foreseeable. Previously, Grok was already caught up in scandals over sexualized deepfakes, and after a wave of criticism in January, some features were made available only to paid subscribers — which, according to the plaintiffs, did not solve the problem but merely monetized it.

"These children may now live their entire lives knowing that such

images exist on the internet," the plaintiffs' attorney described the consequences of the case.

What This Means

These two stories well illustrate that the AI industry is now experiencing a double trust crisis. On one hand, models are penetrating deeper into military and government decisions where the cost of error is measured in lives and political consequences. On the other — even basic consumer products still cannot reliably block obviously dangerous scenarios. For the market, this is a simple signal: the debate around AI is no longer about demos and wow-factor, but about responsibility, control, and real boundaries of what is acceptable.