False AI Intelligence Nearly Triggered US Interception of Chinese Ship

01US military nearly intercepted Chinese ship over false AI-generated intelligence, discovering the report was wrong only before the operation

A false intelligence report nearly prompted the US military to intercept a Chinese ship in the Middle East this spring, amid the war with Iran. The report claimed the vessel was carrying components for a nuclear weapons program, but officials discovered shortly before the planned operation that a chatbot had misidentified its cargo, CNN reported, citing four people familiar with the episode.

The case emerged as the Pentagon expands artificial intelligence across intelligence analysis, strike targeting and logistics. In January, Defense Secretary Pete Hegseth announced an “Artificial Intelligence Acceleration Strategy” intended to make leading AI models available to roughly three million military and civilian personnel across different classification levels. The ship incident shows what changed when that push reached an operational decision: an AI-generated conclusion was converted into a familiar military document and acted upon before its accuracy was established.

The process began when an analyst at a special operations command queried a chatbot about intelligence concerning the ship’s manifest. The underlying reporting originated with US Special Operations Command Pacific. The chatbot combined open-source material with secret signals intelligence held by the government, then wrongly concluded that the vessel was carrying nuclear-program components.

The analyst used AI again to package those findings as a standard intelligence report—the type military officials ordinarily trust—and distributed it. According to CNN’s sources, the document immediately raised alarms and moved through the system without the chatbot’s error being detected.

The military then began planning to intercept the vessel. Two sources said armed US personnel were preparing to board it, while two sources said military aircraft were already airborne. Only shortly before the operation did officials examine the report more closely, discover AI’s role and determine that the cargo had been misidentified. One source described the entire report as false and said it “almost started a war,” reflecting the risk that action against a Chinese vessel could have escalated into armed conflict.

The available account does not identify the cargo, the chatbot, whether it was commercial or government-built, or whether any analyst or reviewer faced consequences. It also does not specify every approval or review stage the report passed. What it does establish is that the standard report format conveyed institutional credibility while the underlying AI conclusion escaped scrutiny until preparations were advanced.

Multiple officials said military and intelligence units use different tools, instructions and safety standards. There is no single standard for verifying AI-generated information, and officials said guidance is also lacking on how human oversight should prevent civilian casualties or friendly-fire incidents as AI use in targeting grows.

02Virginia Moves to Limit Streamlined Data Center Approvals and Bars Officials From Signing Project NDAs

Virginia Gov. Abigail Spanberger has signed an executive order changing how state officials handle data center projects, as rapid development raises concerns about noise, environmental effects and energy prices. Virginia already hosts a cluster described as the data center capital of the world, while some localities have long allowed “by-right approval,” under which developers can build on qualifying land without additional review.

Executive Order 22 immediately bars executive branch officials from signing nondisclosure agreements for data center projects. It also directs the state to expedite noise regulations and review the operation of backup-generation facilities used by data centers. The source does not specify what standards the noise rules will impose or whether the review will restrict when backup generators can run.

Those requirements are separate from Spanberger’s new “Data Center Accountability Framework,” which sets out policy priorities that have not yet been implemented. The framework calls for eliminating some by-right approvals, removing some state subsidies, establishing environmental safeguards and protecting residents from data center-related increases in energy prices.

Ending by-right treatment could give local communities more influence by subjecting qualifying projects to additional approval processes. But the available details do not identify which localities or land categories would be affected, what public participation would be required, or when the change could take effect. They also do not say which subsidies would be removed, which proposals need legislation, or how protections against higher power bills would work.

The executive order also creates an AI task force charged with assessing risks including workforce displacement and data privacy, and examining how existing law could apply to harms involving AI. That work broadens the initiative beyond the physical effects of data centers, although no timetable or specific legal changes were provided.

Some environmental advocates say the measures do not address the full scale of Virginia’s existing development. The Piedmont Environmental Council, a nonprofit that has advocated for a moratorium on new projects, said the approach largely sets standards for new data centers without resolving the cumulative effects of facilities already built or in the development pipeline. The source does not quantify how the measures will affect pending projects or residents’ electricity rates.

03DeepSeek unveils V4.1-Flash, using a smaller KV cache to support million-token context

DeepSeek has introduced V4.1-Flash, a multimodal mixture-of-experts model designed to support contexts of up to one million tokens while reducing the cost of long-running AI agents. Such workloads process large inputs, making both the initial “prefill” computation and the storage of the model’s intermediate context increasingly expensive.

Those intermediate states are held in a key-value, or KV, cache, which grows with the context and consumes memory, storage capacity and transfer bandwidth. DeepSeek’s team says V4.1-Flash tackles these pressures at three points: prefill computation, the global cache kept in high-bandwidth GPU memory, and the persistent cache stored in host memory or on SSDs.

The model has 552 billion backbone parameters, but its Causal Encoder-Decoder architecture activates only part of them for each token. It activates 16 billion parameters per token during decoding, when the model generates its response, and 8 billion during prefill, when it processes the input. The team says the lower prefill activation is intended to improve cost efficiency for input-heavy agent workloads.

For the cache that must remain in GPU high-bandwidth memory, V4.1-Flash combines cross-layer KV-cache reuse in its Compressed Sparse Attention 2 system with FP4 caching, a low-precision representation. According to the paper, these changes reduce the global cache footprint to 890 bytes per token—roughly one-quarter of the corresponding footprint in DeepSeek-V4-Flash.

A separate deployment technique, called SWA Bounded Replay, targets persistent caching rather than the GPU-resident cache. DeepSeek says it cuts the KV-cache footprint that remains on SSDs or in host memory to roughly one-eighth of V4-Flash’s level. The one-quarter and one-eighth comparisons therefore apply to different storage layers, not to a single overall memory reduction.

DeepSeek says V4.1-Flash was pretrained on a multimodal corpus containing 45 trillion tokens and performs substantially better than its V4-Flash baseline despite the smaller cache. Developers can download the released model checkpoints from Hugging Face. The source provides no independent reproduction, service pricing, complete hardware throughput or latency results, or task-level measurements of quality and resource use at the full one-million-token limit.