AI Misidentification Nearly Triggered US-China Military Confrontation

This spring, as the United States was at war with Iran, a piece of intelligence moved through the American military establishment with the weight of official credibility. It claimed that a Chinese ship operating in the Middle East carried components connected to a nuclear weapons program. The report triggered a swift response: the US military began preparing an operation to intercept the vessel. Armed personnel readied themselves to board. Aircraft were already airborne. Then, shortly before the operation was to proceed, someone took a harder look at the intelligence itself.

What they found was a failure at the foundation. A chatbot used by an analyst at US Special Operations Command had misread the ship's manifest. The AI system had taken open-source information and combined it with classified signals intelligence held by the US government, then reached an entirely wrong conclusion about what the cargo contained. The analyst, trusting the machine's output, had then fed those findings back through another AI system to format them as a standard military intelligence report—the kind of document that carries institutional weight and travels through official channels as fact. By the time it circulated, the report bore all the hallmarks of legitimate military intelligence. One source familiar with the episode described it as "entirely false." Another said simply: "It almost started a war."

The incident began in Hawaii, where US Special Operations Command Pacific held intelligence about the ship's manifest. An analyst examined this material and decided to query a chatbot for analysis. The AI system, working from both open and classified sources, produced its misidentification. Rather than flagging uncertainty or requesting human review, the analyst used AI a second time to transform the findings into a formal intelligence product. The report then moved through the system as if it were the product of conventional human analysis. CNN could not determine whether the chatbot was a commercial tool or something built specifically for the US government. A former senior US official offered a telling observation: internal military and intelligence tools were "mostly just copies of the commercial stuff wearing lipstick."

What makes this incident significant extends beyond the single near-miss. The US military and intelligence agencies are rapidly expanding their reliance on AI across a widening range of functions—analyzing vast volumes of intelligence data, selecting targets, repositioning military assets, managing logistics, budgeting, and handling supply chains. Pentagon leadership has argued that AI-assisted decision-making, precisely because it is faster, may prove essential in a future conflict with China or another major power. In January, Defense Secretary Pete Hegseth released the Pentagon's "Artificial Intelligence Acceleration Strategy," explicitly designed to speed up the military's adoption of AI systems. The logic is straightforward: in a high-speed conflict, the side that can process information and act on it more quickly may hold a decisive advantage.

Yet the risks are equally clear. Any US attempt to board or confront a Chinese vessel could have escalated into a broader military confrontation between Washington and Beijing—two nuclear-armed powers with deep strategic competition. An error in judgment, amplified by speed and institutional credibility, could have consequences that extend far beyond the immediate tactical situation. The incident also arrives as governments and technology companies are grappling with a parallel problem: the misuse of AI for weapons development. Anthropic, an AI safety company, recently disclosed that a Yemen-based weapons engineering cell had used its Claude AI model to develop guidance, navigation, and control software for missiles and guided rockets. The group used Claude Code to integrate an open-source autopilot with a flight computer, write control software, adjust settings, build firmware, and run flight simulations—essentially replacing human software engineers with an AI system. Anthropic said safeguards blocked many requests but not all, and the group repeatedly attempted to circumvent restrictions. While there was no evidence the group successfully deployed an operational weapon, they did conduct a test launch of a guided rocket that appeared to fail.

Anthropric identified other concerning conventional-weapons operations as well, including work on drone swarms, torpedo interception systems, and targeting systems for electronic warfare and air-defense suppression. The company also flagged around 35 potentially concerning biological research efforts over a single 30-day period, though it cautioned that it could not determine whether those activities were intended for harm or legitimate scientific research. These disclosures have begun reaching the highest levels of international security discussions. OpenAI CEO Sam Altman is scheduled to brief the UN Security Council on AI and international security, according to Reuters. The meeting reflects a growing recognition among governments that increasingly powerful AI systems require governance frameworks and common safety standards—even as military establishments worldwide accelerate their adoption of the technology.