A real-world AI glitch, not a sci-fi war game, nearly pushed the United States and China toward a direct military confrontation in the Middle East earlier this year, according to multiple reports. An intelligence document built with the help of an AI chatbot incorrectly flagged a Chinese ship as carrying components for a nuclear weapons program, leading US forces to prepare an interception that sources say “almost started a war.”
CNN’s exclusive investigation describes how a Special Operations Command analyst tasked an AI system with examining data on a Chinese vessel operating in the Middle East during the US war with Iran. The chatbot reportedly “fused together” publicly available information with highly classified signals intelligence, then concluded the ship was hauling parts tied to a nuclear weapons program. That assessment was compiled into a formal intelligence report, which moved up the chain of command and triggered planning for US forces to stop and board the vessel.
Sources told CNN that armed US personnel were on the verge of launching the operation when officials took a harder look at the intelligence product. Only then did they discover the report had been prepared with significant input from an AI chatbot, which had simply gotten the cargo wrong. One source described the document as “entirely false” but warned that it still “almost started a war,” since any US move against a Chinese vessel—especially amid a live conflict involving Iran—could have spiraled into a wider clash between nuclear-armed powers.
Outlets from tech-focused Engadget to international papers have echoed CNN’s account, underscoring that the core failure was not a rogue robot but a trusted analyst leaning too heavily on a probabilistic text generator in a high-stakes environment. The AI system wasn’t some bespoke military targeting engine; it was a chatbot tasked with synthesizing different streams of data and presenting a narrative, a job large language models excel at stylistically even when the underlying facts are wrong. By confidently stitching together open-source material and secret signals intelligence into a compelling but inaccurate story, the bot turned a mix of real data and bad inference into something that looked like actionable evidence.
Officials ultimately aborted the boarding operation once the AI’s role and the misidentified cargo became clear, and the ship continued on its way. The episode, however, has become an instant cautionary tale about the Pentagon’s rapid push to bolt AI onto existing intelligence and command workflows. US defense leaders have been loudly evangelizing AI as a way to process sensor data at machine speed and outpace rivals in future conflicts, but this incident shows how a single hallucinated conclusion can propagate through a system that assumes the inputs are solid.
For geeks and tech watchers used to treating “AI alignment” and “Skynet scenarios” as abstractions, this is the kind of story that snaps the stakes into focus: a misfire from what amounts to an extremely fancy autocomplete nearly triggered live military action between superpowers. It exposes a core tension in AI-for-war: the same tools that promise faster analysis and slicker briefings can also manufacture reality with enough confidence to fool human gatekeepers, especially under time pressure. As militaries race to plug chatbots and pattern-recognition engines into everything from threat assessments to targeting, the near-miss in the Middle East will likely be cited for years as Exhibit A in the argument that humans need more than a cursory “sanity check” before letting machine-written intel nudge them toward the next crisis.
