Grok Pentagon Iran targeting revealed in sworn court filing over deadly strikes

A sworn declaration buried in an environmental lawsuit over a Memphis data center has done what no Pentagon press release did: it confirmed that Elon Musk's xAI put an artificial intelligence model at the center of one of the deadliest American air campaigns in years. The court filing, submitted by the Pentagon's chief digital and artificial intelligence officer, states that the "Grok Gov Model" helped U.S. forces deploy more than 2,000 munitions against 2,000 distinct targets inside Iran in a span of just 96 hours.

The revelation about Grok Pentagon Iran targeting has thrown a harsh light on how quickly commercial AI has moved from consumer chatbots to the machinery of lethal war planning, and on how little the public knew about it until a routine discovery fight forced the details into the open. More than 120 members of Congress are now demanding answers, and two senators have introduced legislation aimed squarely at the practice.

The court filing that exposed Grok's combat role

Cameron Stanley, the Pentagon's chief digital and artificial intelligence officer, disclosed the AI model's role in a sworn court declaration. According to that declaration, the Grok Gov Model helped U.S. forces deploy over 2,000 munitions to 2,000 distinct targets within 96 hours during an offensive against Iran that the military labeled "Operation Epic Fury."

What makes the disclosure extraordinary is where it surfaced. It did not come from a Pentagon briefing, a congressional hearing, or a Defense Department fact sheet. It came through a federal court filing in a Mississippi environmental lawsuit brought by the NAACP over pollution from xAI's Colossus 2 data center near Memphis. In other words, the acknowledgment that a Musk-owned AI helped select thousands of strike targets emerged as a byproduct of litigation about air quality and industrial emissions.

That accidental route to disclosure is now part of the story. Lawmakers who have spent months pressing for information about xAI's military work learned key operational details not from the executive branch but from a plaintiff's discovery in an unrelated case. The gap between what the Pentagon told Congress and what it told a Mississippi court has become its own line of inquiry.

How Grok was embedded inside Project Maven

Grok Gov did not operate as a standalone tool. It was embedded inside the Pentagon's Maven Smart System, the software layer at the heart of Project Maven, the military's flagship effort to use machine learning to sift through surveillance and intelligence data. Inside that system, according to the disclosures, Grok processed both classified and open source intelligence and flagged potential targets for human planners.

The practical effect, officials have said, was speed. Grok compressed planning timelines that once took days down to hours. In a campaign that struck 2,000 targets in less than four days, that acceleration was not incidental; it was the point. The model's ability to ingest streams of imagery and reporting and surface candidate targets is precisely what allowed the tempo of Operation Epic Fury.

Speed, however, cuts both ways. Compressing the target development cycle from days to hours leaves less time for the layered human review that has traditionally guarded against errors in intelligence, identification, and legal vetting. Critics of the arrangement argue that the same feature marketed as a battlefield advantage is also the feature most likely to let a mistake through before anyone catches it.

Grok Pentagon Iran targeting

Scrutiny of the Grok Gov Model's role sharpened dramatically because of one strike in particular. In the early hours of Feb. 28, 2026, a munition hit the Shajareh Tayyebeh girls' school in Minab, Iran, killing more than 170 people, most of them children. It is one of the deadliest single incidents attributed to the campaign, and it is the case now driving the broader debate over Grok Pentagon Iran targeting on Capitol Hill.

NBC News, citing four sources, reported that preliminary information shows a U.S. munition was likely responsible for the school strike, and that outdated intelligence may have contributed to selecting the target. That detail, outdated intelligence feeding an automated targeting pipeline, is the nightmare scenario that AI safety advocates had warned about, and it now sits at the center of the congressional response.

More than 120 Democratic members of Congress sent a letter to Defense Secretary Pete Hegseth on Thursday, March 12, 2026, obtained by NBC News, asking whether the Pentagon would investigate the school strike as a possible war crime. The letter reframed a technical debate about software procurement into a direct question about accountability for civilian deaths, and it tied that question explicitly to the role automation played in choosing targets.

Warren's letter and the guardrails question

Sen. Elizabeth Warren, D Mass., a member of the Senate Armed Services Committee, escalated the pressure on March 16, 2026, with a letter to Hegseth demanding documentation of xAI's safety assurances. Warren warned that "Grok's apparent lack of adequate guardrails could pose serious risks to the safety of U.S. military personnel," turning the model's public track record for erratic and unfiltered outputs into a national security argument.

Warren's framing is notable because it does not rest solely on the risk to civilians abroad. By emphasizing danger to American service members, she broadened the coalition that might care about the issue and pressed the Pentagon on whether it independently validated xAI's guardrails or simply accepted the company's word.

The letter also demanded documentation, not reassurance. Warren asked for the paper trail: the safety assurances xAI provided, the standards the Pentagon applied, and the evidence that anyone tested Grok's behavior before wiring it into a system that flags people for lethal strikes. As of this writing, the extent of that documentation remains a subject of the ongoing dispute.

The February deal that put xAI on classified networks

The pathway to Grok's battlefield role runs through a deal struck weeks before Operation Epic Fury. In late February 2026, the Pentagon and xAI reached an agreement, first reported by Axios, allowing xAI's systems to be used on classified networks. That access is what made it possible to feed Grok the classified intelligence it processed inside the Maven Smart System.

The timing matters. The deal came amid the Pentagon's rupture with a rival AI company, Anthropic, which had refused to drop restrictions on domestic surveillance and autonomous weapons use. Where Anthropic held its lines, xAI agreed to an "all lawful use" standard, and that willingness helped clear its path onto sensitive military systems.

The commercial relationship predated the classified network deal. In July 2025, xAI received a contract worth up to $200 million from the Pentagon's Chief Digital and Artificial Intelligence Office to develop AI applications for the Defense Department. The February agreement built on that foundation, moving xAI from developing applications to operating inside the networks where the most sensitive targeting work happens.

GenAI.mil and Grok's reach across 3 million users

The military's embrace of Grok extends well beyond a single campaign. The Department of War, as the Pentagon has been rebranded, separately announced that Grok would deploy across systems serving roughly 3 million military and civilian personnel under a platform called GenAI.mil. That is a scale of adoption far larger than any one operation.

Central to that arrangement is the same "all lawful use" standard that distinguished xAI from its competitors. xAI agreed to a use standard that Anthropic had declined for its Claude model, and that difference in corporate policy translated directly into different levels of military access. The company most willing to accept broad terms became the company most deeply embedded.

The GenAI.mil rollout means the questions raised by the Iran campaign are not confined to one theater or one mission. If Grok flags targets, drafts analyses, and processes intelligence for millions of users across the department, then the guardrails debate is not about a pilot program. It concerns the default AI layer of the American military.

Gillibrand's bill and the fight over autonomous targeting

Congress has begun translating alarm into draft law. Sen. Kirsten Gillibrand, D N.Y., introduced the "Secure and Accountable Military AI Act," legislation that would set hard limits on how the Pentagon uses artificial intelligence in combat. The bill would ban AI from targeting or launching nuclear weapons and ban the development of fully autonomous weapons.

Crucially for the Iran case, the bill would also require senior level written approval before the military deploys "high consequence" AI for lethal targeting support. That provision speaks directly to the compressed timelines that Grok enabled: it would force a named, accountable official to sign off in writing before an AI system is turned loose on the target development cycle.

Whether the bill can pass in a divided Congress is uncertain, but its introduction changes the terms of the debate. It establishes a concrete legislative benchmark against which the Pentagon's current practices can be measured, and it puts every lawmaker on record about where the line between human judgment and machine speed should fall.

The accountability gap the Pentagon now faces

The through line connecting every strand of this story is a gap between capability and accountability. The Pentagon acquired a tool that can select 2,000 targets in 96 hours before it had settled the public questions about oversight, guardrails, and responsibility for error. The disclosure that surfaced through Grok Pentagon Iran targeting has now forced those questions into the open all at once.

Several concrete threads will determine how this resolves. The Pentagon has been asked whether it will investigate the Minab school strike as a possible war crime. Warren is awaiting the documentation of xAI's safety assurances. Gillibrand's bill offers a template for written approval requirements. And the NAACP's environmental lawsuit, the case that started it all, continues to move through federal court in Mississippi.

For now, the record consists of a sworn declaration, a body count, and a set of unanswered congressional letters. The Pentagon chose the AI vendor most willing to accept broad terms, embedded its model in the system that helped pick thousands of targets, and did not disclose the arrangement until a court compelled it. The reckoning over that sequence of choices is only beginning.