The Kill Chain Report: Pentagon Review Blames AI Overreliance for the Strike That Killed 123 Children in Minab

On February 28, 2026, two Tomahawk cruise missiles struck the Shajarah Tayyebeh Elementary School in Minab, a port city of roughly 76,000 people on Iran’s southern coast near the Strait of Hormuz. The first missile hit shortly after 11:00 local time, as teachers were sending children home early amid reports of US attacks. A second arrived minutes later, collapsing part of the roof onto a prayer room full of children. More than 150 people were killed, at least 123 of them children. It was, by child casualties, the deadliest US military targeting error of the 21st century.

Now we know how it happened. A Bloomberg investigation published September 18 — built on interviews with more than two dozen current and former officials involved in an internal Pentagon review ordered by CENTCOM chief Admiral Brad Cooper — reconstructs the “kill chain” that turned an elementary school into a target. And at the center of that chain sits a familiar Silicon Valley name: Palantir’s Maven Smart System, the AI platform that has become the Pentagon’s favored targeting tool, running workflows built on Anthropic’s Claude.

What the review actually found

Investigators described not one error but a chain of preventable failures:

Stale data, ignored signals. The Minab site had been classified as an Islamic Revolutionary Guard Corps facility for years — correctly, once. The school had been built nearly a decade earlier on land that once belonged to an IRGC naval unit. Commercial satellite imagery showed the separation was completed by 2017; a 2018 image showed brightly painted walls, a football pitch, and markings for children’s activities. One intelligence analyst noticed the changes as early as 2019 — but recorded the observations in a system that was never linked to the main targeting database. The primary database kept saying: IRGC facility. Targeting officials, per an earlier New York Times report, worked with satellite imagery up to seven years old.

A compressed timetable. The Trump administration ordered a major aerial assault in the final days of February, and more than 1,000 Iranian targets were struck in the first 24 hours. The pace left planners with a fraction of the usual time to finalize targets. Officials told Bloomberg that Maven — which fuses more than 150 data sources into a single operating picture — compressed work that previously took hours into minutes. Speed, in this case, didn’t just enable the campaign; it stripped out the friction where human doubt used to live.

Gutted civilian-protection teams. Staffing across the Pentagon’s civilian-harm mitigation teams was cut by roughly 90%, leaving fewer than 20 personnel. At Central Command, the team fell from ten people to one. No civilian-harm specialist reviewed the Minab site before the strike — a review that was routine in prior years, though not mandatory. Defense Secretary Pete Hegseth had framed the cuts as shedding “Biden-era, non-lethal programs,” and days after the strike celebrated “the most lethal and precise air power campaign in history” with “no stupid rules of engagement.”

Overreliance on AI. The review’s most consequential finding: some Central Command personnel relied too heavily on Maven, expecting the system to flag outdated information or inconsistencies in target intelligence — capabilities the system was never trained to provide. Officials quoted in earlier reporting describe Maven, with Claude-based workflows, semi-autonomously ranking targets for the opening salvos. The machine didn’t hallucinate a school into existence; it faithfully fused stale, siloed data into a confident picture, and harried humans outsourced their skepticism to it.

Palantir’s defense — and its quiet fix

Palantir’s response is a study in contract-line lawyering: the company says it is “not responsible for the underlying data nor identifying intelligence deficiencies,” and that there is no evidence its software was at fault. Two people familiar with its Pentagon contracts note the government remains primarily responsible for data quality going in.

Yet according to a person familiar with the matter, Palantir has since added capabilities to Maven that “re-review underlying intelligence to identify factors that would disqualify a target and flag inconsistencies and inaccuracies that human review may have missed” — and the new tools have already surfaced anomalies. A feature shipped after 123 children died is an implicit admission that the gap was real.

The reckoning accelerates

The report landed in the middle of an already-sensitive week. On Thursday, the UN’s Independent International Fact-Finding Mission on Iran concluded there were “reasonable grounds” to believe the Minab strike, and another attack the same day, amounted to war crimes, finding the US “failed in its obligation to do everything feasible to verify” the target — a failure the mission said “went beyond negligence.” The US has not publicly accepted responsibility; President Trump has suggested Tehran bore the blame, telling Fox News in July, “I don’t think anybody’s going to ever be able to say what happened there.”

On Saturday, Senators Mark Warner, Jack Reed, and Chris Coons sent a letter to Hegseth and Director of National Intelligence Jay Clayton demanding “immediate investigation by relevant Inspectors General” into AI-enabled targeting failures — citing both Minab and CNN’s report that an AI-generated, “entirely false” intelligence report about a Chinese cargo ship nearly triggered a US interception operation. The senators warn that “AI targeting platforms have repeatedly generated spurious outputs (based on outdated input data or outright hallucination),” and demand IG access to “any additional instances that may have so far not been publicly reported.”

There is also an awkward corporate footnote: CNN reports the Defense Department has sparred with Anthropic over the lab’s refusal to strip safety guardrails that prevent Claude’s use in autonomous weapons and mass surveillance. Anthropic’s position — that its models aren’t reliable enough for those purposes — reads, this week, less like caution and more like prophecy.

The real lesson

The temptation is to file Minab under “AI kills children.” The review says something subtler and more damning: the failure was sociotechnical. Seven-year-old imagery, siloed databases, a 90% staffing cut, a compressed timetable, and an AI system trusted to do a job it was never designed for — each layer converting the last layer’s output into confident-seeming ground truth. Automation bias is not a hypothetical from a safety paper; it is now the documented mechanism behind the deadliest targeting error of a generation.

The Pentagon’s full report remains unreleased, though officials say it has been effectively complete for months. Until it sees daylight, the accountability chain — who saw the 2019 analyst note, who approved the target, who decided civilian-harm reviews were optional — stays inside the building. The Senate letter gives inspectors general a lever to pry it open. Whether they use it will tell us whether “the United States does not target civilians” is a commitment or a caption.