A military AI agent cut World Cup crisis reports from 20 minutes to 90 seconds; for now the figure is only the Pentagon's word

By Federal News Network · November 9, 2026.

An AI tool of military origin helped thwart a bomb threat during the 2026 FIFA World Cup, according to Capt. Ryan Hetrick, who leads the Pentagon's Generative AI Task Force (GenAI Task Force). He made the remarks on November 6 at an AI event in Washington. The incident had not been made public until now.

Here is how Hetrick tells it: on June 18, during the Switzerland–Bosnia and Herzegovina match at SoFi Stadium, near Los Angeles, a threat was triggered. Three days earlier, personnel from the Department of Defense, the FBI and the Coast Guard had met at the White House to design an automated agent to draft "spot reports": short, urgent reports on fast-changing situations.

The workflow goes like this. When there is a security incident, law enforcement on the ground calls the White House. The World Cup task force drafts the report and sends it up the chain of command to coordinate the response. That process usually takes about 20 minutes. The agent does it in 90 seconds, which, according to Hetrick, gave "18 extra minutes" to coordinate on the ground. "That bomb threat was deterred," he said.

It is worth being careful about what we actually know. Federal News Network notes that spokespeople for the White House, the FBI and local police did not immediately respond to its requests for comment. In other words, the 20-minutes-versus-90-seconds figure and the causal link to the threat rest, for now, on the account of an official who is also an avowed evangelist for these tools. Nor are there details on what the threat was or how it was neutralized.

For context, the tournament drew more than six million spectators across more than 100 matches and 16 venues in the U.S., Canada and Mexico. FBI Director Kash Patel said it was the highest level of support the agency has ever provided to a global sporting event. According to the outlet, two other people were charged over a separate bomb threat at MetLife Stadium in New Jersey.

Hetrick offered another example: at the 101st Airborne Division at Fort Campbell (Kentucky), he built a tool that cut a cyber operations workflow from 100 hours to 15. He says it is now available to every cyber officer in the department. His approach: "We don't do policy shops or PowerPoints. We go to the edge and build with the warfighter."

**Our take.** My thesis: this is the easiest use of agentic AI to defend, and that is precisely why it deserves rigorous scrutiny. Drafting a structured report from a phone call is a bounded task, with a human making the decision and a clear metric: time. It replaces no one's judgment; it compresses bureaucracy. Here the AI doesn't "detect" the bomb—it gets information to the decision-maker faster.

That pattern matches what we see in other sectors: the value lies in the time freed up for what matters, not in the volume of automation. And differentiation, as we have argued before, lies in workflow and permissions design, not in the model. The fact that it was designed just three days before the incident, with Defense, the FBI and the Coast Guard in the same room, is the most transferable lesson.

The risks are real too. A crisis report generated in 90 seconds can spread an error through the entire chain of command just as quickly. The question Hetrick leaves unanswered in what has been published is what human review takes place before the report is sent, and what happens when the agent gets it wrong. In a security setting, speed without verification is a risk, not an advantage.

In the short term, this accelerates the adoption of agents in defense and public safety, with the accountability questions that come with it. In the long term, the same logic—getting paperwork out of the way so people can focus their attention on judgment—is what could free up human time in medicine, justice or emergency services. But to fully believe it, we need what is still missing: independent confirmation and data on errors.