False AI Intel Almost Sent US Forces Onto a Chinese Ship

A chatbot-built US intelligence report this spring sent planes up over a Chinese ship said to be carrying nuclear-weapons parts. Armed crews were preparing to board. Officials stopped the operation only after they traced the finding to an AI tool that had misidentified the cargo, four people familiar with the episode said.

One of them called the report entirely false. The same person said it almost started a war. Any boarding of a Chinese vessel in the middle of the Iran campaign could have spiraled into a fight between the two countries.

What Happened When a Chatbot Misread a Chinese Ship’s Cargo?

The paper went out across the US military in spring 2026, during the war with Iran that began on Feb. 28. It claimed a Chinese ship in the Middle East was moving parts of a nuclear weapons program. Two of the four people said armed service members were getting ready to go aboard. Military planes were already in the air, according to one of those people and another person who knows the incident.

The write-up came from a special operations command analyst. The underlying notes on the ship’s manifest originated with Special Operations Command Pacific in Hawaii, the SOF component for Indo-Pacific Command, which covers 51% of the globe and 60% of the world’s population from Camp H.M. Smith. The analyst asked a chatbot about that reporting. The bot mixed open-source material with secret signals intelligence in government files and named the cargo wrong. What the ship was actually carrying has not been disclosed.

It was not clear whether the chatbot was a commercial product or a government build. A former senior US official who knows the systems used by military and intelligence analysts said the in-house versions are mostly copies of the commercial ones wearing lipstick. The analyst then used AI a second time to wrap the finding in a standard intelligence report, the format commanders treat as finished work, and sent it out.

Only just before the planned operation did officials dig into how the paper had been made. The Pentagon and the Hawaii command did not comment on the episode.

WHAT WE KNOW

• The claim: A Chinese ship in the Middle East was said to be carrying nuclear-weapons parts.
• The response: Armed crews prepared to board and military planes were in the air.
• The break: Officials stopped the operation after tracing the finding to a chatbot that named the cargo wrong.
• The product: The analyst used AI once to reach the conclusion and again to package it as a standard report.

WHAT IS UNCONFIRMED

• The tool: No one has said whether the chatbot was commercial or a government copy.
• The cargo: The actual load on the ship has not been identified.
• The date: The exact day in spring 2026 has not been given.
• The aftermath: There is no public account of discipline, a stand-down, or a new verification rule.

The near-miss was not a weapon that fired itself. It was a trusted document with a model inside it, moving through a wartime chain of command built to act fast.

Hegseth Told Three Million People to Make AI Their Teammate

That paper landed after Secretary of War Pete Hegseth spent the winter ordering the department to go faster, not slower. On Jan. 9, 2026, he signed a memo titled Artificial Intelligence Strategy for the Department of War. The department rolled it out on Jan. 12 as an “AI-first” fighting force, with seven Pace-Setting Projects under three headings: warfighting, intelligence, and enterprise operations.

The memo’s enterprise aim was putting frontier models in three million hands, at every classification level. Hegseth said the department would unleash experimentation, cut bureaucratic barriers, focus its spending, and show it could lead in military AI.

I expect every member of the department to log in, learn it and incorporate it into your workflows immediately. AI should be in your battle rhythm every single day; it should be your teammate. By mastering this tool, we will outpace our adversaries.

Pete Hegseth, Secretary of War, message introducing GenAI.mil

GenAI.mil itself went live on Dec. 9, 2025, first with Google’s Gemini for Government, cleared for controlled unclassified information at Impact Level 5. On Aug. 31, 2026, the War Department added OpenAI’s ChatGPT Mil and Starshield AI’s Grok for Government on the same platform. An Aug. 31 release put unique users at more than 1.7 million out of more than three million military and civilian personnel.

THE SEVEN PACE-SETTING PROJECTS

Open Arsenal is the intelligence line that states the tempo in plain language: hours, not years. Agent Network, launched on June 25 with Pacific Command among the partners, is the warfighting line that puts agents on the kill chain. Neither project is a verification standard for a chatbot that has just invented a cargo list.

The Report Looked Like Any Other Trusted Product

The spring incident had two AI passes, not one. The first pass was the finding. The second was the packaging. That second pass is what made the error travel.

HOW THE FALSE REPORT WAS BUILT

1. The raw file: Manifest reporting on a Chinese ship in the Middle East originates with SOCPAC in Hawaii.
2. The query: A special operations command analyst asks a chatbot about that reporting.
3. The fusion: The bot mixes open-source intelligence with secret signals intelligence and names the cargo as nuclear-weapons parts.
4. The write-up: The analyst uses AI again to drop the finding into a standard intelligence report.
5. The send: The report goes out in the format military officials already trust.
6. The spin-up: Forces plan an intercept; armed crews prepare to board; planes go up.
7. The catch: Officials dig into the paper just before the operation and find the chatbot error.

Hallucinations of this kind have not been a one-off since these tools spread across government, one of the people said. Older intelligence officials, including some who support using AI, say the tools have pushed analysts to produce and send work faster, which opens the door to mistakes. Younger analysts grew up on the same products and are more likely to trust them without a second look, several people said.

AI allows you to get to a bad idea faster.

A US official familiar with the episode

The model did not sneak onto a target list by itself. A person asked it a question, accepted the answer, dressed that answer as finished intelligence, and a wartime staff treated the finished look as a reason to move. The only brake was a later official who decided to check the work.

Maven Already Packs a Day of War Into a Thousand Targets

The chatbot report sat beside a much larger targeting stack that was already compressing the same loop. Palantir is the prime software integrator for Maven Smart System, the department’s flagship AI targeting platform. A June analysis by the Center for Strategic and International Studies found the United States used Maven to help strike more than 1,000 targets in 24 hours at the start of the Iran war, a tenfold jump over the pre-Maven era. CSIS put the user base at roughly 80,000 as of that June paper. The Maven contract ceiling surpassed $1 billion in May 2025.

MAVEN, IN PUBLIC FIGURES

• First-day tempo: More than 1,000 targets in the first 24 hours of the Iran war, per CSIS.
• Users: Roughly 80,000 on the platform, CSIS estimated in June 2026.
• Money: The Maven contract ceiling surpassed $1 billion in May 2025.
• Staffing comparison: In Scarlet Dragon, about 20 people matched a targeting cell of more than 2,000 from the Iraq war.

Cameron Stanley, then the department’s chief digital and AI officer, described the change in a March presentation. Work that once jumped across eight or nine systems now sits in one.

We’ve gone from identifying a target to now coming up with a course of action, to now actioning that target. All from one system. This is revolutionary.

Cameron Stanley, Chief Digital and AI Officer, March 2026 presentation

On June 25 the War Department launched Agent Network, the second Pace-Setting Project, with Pacific Command, Southern Command, and European Command as partners. The release said the tools would turn intelligence into informed options for commanders and shrink the time from a critical development to an act. Palantir’s Maven work is named as a foundation. The point of the project is speed.

A chatbot that misreads a manifest is not Maven picking a yellow box on a drone feed. It is a cousin of the same bet: fuse more sources, write faster, send the product up the chain in a form people already trust. When the product is a strike option, a bad box can kill the wrong people. When the product is a boarding order on a Chinese ship, a bad sentence can start a different war.

Young Analysts Trust the Tools Older Officers Still Check

The department’s AI push is not one system with one safety rule. Different offices use different tools under different orders, multiple US officials said. There is no single standard for how the United States checks what these tools generate. Reliability varies with the shop that fielded the model.

Hegseth has told Congress that humans, not AI agents, make the lethal call. “We follow the law and humans make decisions,” he told the Senate Armed Services Committee in May. “AI is not making lethal decisions.” That sentence answers a question this episode does not ask. No one claims a model ordered the boarding. The claim is that a model wrote the reason to board, in the voice of an intelligence report, and people almost did it.

Another person familiar with current military policy said AI in targeting is ramping up and that there is no real guidance for how a human in the loop will prevent civilian deaths or friendly fire. A human who sees a finished page, under pressure to move in a war, is not the same as a human who rebuilds the file from the raw take. The spring catch happened because someone did the slower job. The January memo told the force to stop waiting on that kind of friction.

The loud argument in Washington this year has been whether frontier models might someday slip human control. This case is smaller and closer. A non-deterministic tool was asked about a ship, fused two kinds of intelligence, and emitted a confident wrong answer. Asked twice, it might have emitted a different wrong answer. That is a known property of these models. It is a poor property for a document that can put armed crews on a foreign deck.

The Human-Judgment Rule Covers Weapons, Not Invented Cargo

The written rule most often cited for this problem is still aimed at weapons, not at reports. DoD Directive 3000.09, updated on Jan. 25, 2023, says autonomous and semi-autonomous weapon systems will be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force. A chatbot that labels cargo, then drafts the cable, is not a weapon system under that text. The boarding party would have been the force. The error sat one layer upstream, where the directive is quiet.

On June 5, 2026, the White House issued National Security Presidential Memorandum 11 on AI in the national security enterprise. It told the secretary of war to issue an update to Directive 3000.09 within 90 days, which ran to Sept. 3. The memo also told the national security enterprise to put frontier models in professionals’ hands without delay, and to run testing, evaluation, validation, and verification so those systems stay reliable and controllable. Those two instructions sit in the same document. The spring episode is what they look like when the first one outruns the second.

THE NINE-MONTH AI PUSH

1. December 9, 2025: GenAI.mil launches with Gemini for Government at IL-5.
2. January 9, 2026: Hegseth signs the AI Acceleration Strategy memo.
3. January 12, 2026: The War Department announces the seven Pace-Setting Projects and an “AI-first” force.
4. February 28, 2026: The war with Iran begins; Maven is used in the opening surge.
5. Spring 2026: A chatbot-built report almost sends US forces onto a Chinese ship.
6. June 5, 2026: NSPM-11 orders a 90-day update to Directive 3000.09.
7. June 25, 2026: Agent Network launches to compress intelligence into commander options.
8. August 31, 2026: ChatGPT Mil and Grok for Government join GenAI.mil; unique users top 1.7 million.

The analyst who queried the ship’s manifest has not been named. The chatbot has not been named. The cargo has not been named. What has been named is the policy that told three million people to treat these models as teammates, and the absence of a common test for the sentences those teammates write. The boarding did not happen because a person, at the last moment, opened the file again.