AI hallucination nearly led the US military to board a Chinese ship

What happened

In spring 2026, during the war between the US and Iran, the US military came close to intercepting and boarding a Chinese ship in the Middle East. The trigger was an intelligence report produced with the help of an AI chatbot. It claimed the freighter was carrying components for a nuclear weapons program. That was false. The incident came to light on September 18, 2026, in an exclusive CNN report based on four people familiar with the matter. According to two of those sources, armed troops were already preparing to board and military aircraft had already taken off. Only when officials took a closer look at the underlying information did the error surface, and the operation was called off.

"Nearly started a war." – a CNN source

Another source called the report "completely wrong." CNN couldn't determine what the ship was actually carrying. Neither the Pentagon nor U.S. Special Operations Command Pacific (SOCPAC), headquartered in Hawaii, responded to requests for comment. Several specialist outlets and commentators have since used the case to make a broader critique, including the Spanish site Moncloa and the US security outlet The Cipher Brief.

How a chatbot answer became a "finished" intelligence report

CNN, Tech Times and gCaptain all describe the process as two steps. First, a SOCPAC analyst asked a chatbot to analyze information about the ship's cargo manifest. The program was supposed to combine publicly available sources (open-source intelligence, or OSINT) with classified signals intelligence (SIGINT, meaning intercepted communications). The chatbot concluded that nuclear weapons parts were on board.

In the second step, the analyst didn't have that result independently verified. Instead, he used the same chatbot again to turn it into a report in standard military format. That document then moved up the chain of command. According to gCaptain, the familiar format made the flawed analysis look like an ordinary intelligence product. Tech Times sums up the problem: there was a human involved at every step, but that human was only writing prompts and passing results along, not checking them.

The mistake is a textbook hallucination: a language model invents facts and states them with the same confidence as accurate information. Tech Times points to an OpenAI study from September 2025. It found that language models hallucinate for structural reasons, because training and evaluation reward guessing and penalize admitting uncertainty. Moncloa puts it like this:

"A model doesn't lie: it fills gaps. And in intelligence, a gap filled with grammatical certainty is more dangerous than an openly acknowledged void."

The columnist recalls intelligence directive ICD 203. It was created after the debacle over Iraq's supposed weapons of mass destruction and requires analysts to document source quality and clearly separate knowledge from assumption. A generative model can't make that distinction. As a historical parallel, Moncloa doesn't cite Stuxnet but "Curveball": the Iraqi defector whose invented mobile bioweapons lab made it all the way to the UN Security Council. The difference, Moncloa argues, is that back then the deception came from a person, while today it can come from a system nobody has checked.

Speed over oversight: the Pentagon's AI strategy

The incident comes as the US Department of Defense is pushing to adopt AI as fast as possible. On January 12, 2026, Defense Secretary Pete Hegseth unveiled an "AI Acceleration Strategy." The goal is an "AI-first" force. The department wants to "unleash experimentation and eliminate bureaucratic barriers" and give all three million civilian and military employees access to leading AI models at every classification level. Under the strategy, new models are to be deployed no later than 30 days after their public release. The strategy also includes a project called "Agent Network": AI agents are meant to support battle management and the so-called kill chain, the military targeting sequence of find, fix, track, target, engage, assess. In his speech, Hegseth counted testing and authorizations to operate among the obstacles that should be treated as "operational risks." According to Defense One, AI ethics doesn't appear in the strategy at all.

The numbers for the GenAI.mil platform show how fast this is moving:

One CNN source sums up the risk in a single sentence: "With AI, you get to a bad idea faster." When AI speeds up the intelligence steps at the start of the kill chain, a false conclusion also travels up faster, leaving less time for human oversight.

The testing and verification gap

Several US officials told CNN that different service branches and agencies use different AI tools under different rules and security standards. There's "no uniform set of standards" for checking AI outputs. There's no mandatory human review before AI analyses are passed on. A former senior official described the government's internal AI tools as "mostly copies of the commercial stuff with lipstick on." One source described the case as part of a "trend in the US military."

The Cipher Brief reviewed eight public announcements about GenAI.mil from December 2025 to September 2026. None of them says what performance tests the models went through or who approved their rollout. Instead, they all point to "Impact Level 5" accreditation. But that only covers the IT infrastructure the models run on, not what the models actually do. A July 2025 guidance document from the Pentagon CIO says so explicitly: AI models don't need an authorization to operate, only the infrastructure does. Adding a model to an already approved system generally doesn't require a new authorization. Performance expectations for the AI system are explicitly excluded.

At the same time, the department has cut the staff of its independent operational test office in half. In April, the US Government Accountability Office (GAO) found that there are still no "clearly defined, universal AI tests." In June, it warned that the cuts raise the risk of fielding systems with undocumented weaknesses. The Pentagon's own testers admitted back in 2025 that testing generative AI is "still in its infancy."

The Cipher Brief authors propose certifying the system rather than the ever-changing model. That would require a written specification defining what correct output looks like. An independent body would automatically rerun it after every model change. Such a test could check, for example, whether the cited evidence actually supports a claim about a ship's cargo. Worth knowing: one of the authors runs Jaxon.AI, a company that sells exactly this kind of verification tooling, including to the Air Force and Space Force. The authors disclose this conflict of interest themselves.

Political reactions and diplomatic context

Democratic Senators Mark Warner (vice chair of the Intelligence Committee), Jack Reed (ranking Democrat on the Armed Services Committee) and Chris Coons wrote to Hegseth and Director of National Intelligence Jay Clayton. They're calling for an "immediate investigation by the relevant inspectors general" with as much transparency as possible. They're concerned that AI adoption and "experimentation" have taken priority over effective oversight. They also want full access to both cases this year in which media reported significant errors in AI-assisted targeting processes, and to any other cases that haven't yet come to light.

The CNN report came out a few days before a summit between Donald Trump and Xi Jinping at the White House on September 24. According to Tech Times, AI governance was on the agenda there too. In 2025, the think tank SIPRI warned that military AI can shorten decision times in crises and increase the risk of miscalculation. Jake Steckler of the Centre for the Governance of AI, a former US Army officer, told TechCrunch that soldiers need to understand the uncertainty of language models, especially in decisions about the use of force. But he also warned that putting speed above everything else risks incidents that destroy trust and could end up slowing adoption.

As a counterpoint, Moncloa cites an incident from July 2026: during an internal cybersecurity evaluation at OpenAI, several models bypassed their isolation controls and broke into parts of the research infrastructure and Hugging Face's systems. According to OpenAI, the agents shared attack paths via a shared message board. So even the developers can't reliably predict how their models will behave.

Our take

This is probably the most concrete documented case yet of an AI hallucination nearly triggering an armed confrontation between two nuclear powers. The problem was less any single model than a process chain with no safeguards: the same chatbot did the analysis and the formatting, the official format lent it false credibility, and in the end a person stopped it, not a control mechanism. Moncloa stresses that there's no evidence of deliberate manipulation by China. The error originated inside the US system itself.

A lot remains unknown. We don't know which chatbot was involved. Also unclear are the actual cargo and how close the operation really came to a boarding. All the details come from anonymous sources, and the Pentagon isn't talking. Congress has set June 2027 as the deadline for legally required standards. Whether the investigation happens, and what it finds, is also an open question. For allies and mid-sized intelligence services that rely on US findings, Moncloa raises one more question: how do they check whether a report that looks completely official is based on a made-up AI answer?