RealityRewi.red

Malaysia caught 30 smuggled AI servers. OpenAI's answers reportedly left by API

Customs in Kuala Lumpur found $12 million in AI hardware labeled as ordinary electronics. The same day, a report said two Chinese labs trained on GPT-6 Astra's output instead.

Reality Admin//Oct 28, 2026//3 MIN READ
Malaysia caught 30 smuggled AI servers. OpenAI's answers reportedly left by API

On October 15, customs officers at Kuala Lumpur International Airport opened a shipment in the airport's free trade zone. The paperwork said electronic components. Inside were 30 servers packed with advanced AI chips, worth 49.5 million ringgit, or about $12.1 million.

Malaysia announced the seizure on Wednesday. The airport's customs director, Aspalila Ag Tuah, said the servers were "falsely declared as ordinary electronic goods." The shipment had come from one Asian country and was meant to go on to another. Officials named neither, and didn't say who made the chips. The Malaysian company handling the shipment has been called in to help the investigation, and no arrests have been announced, according to the state news agency Bernama.

It isn't a one-off. On June 5, customs at the same airport found 72 servers with AI chips, worth about $13 million, also declared as computer parts. Aspalila said investigators link the same network to at least five cases worth roughly $25 million in total.

Why Malaysia

The U.S. bars exports of Nvidia's top chips and other advanced American AI hardware to China. For a while, a well-worn way around that was to ship to a friendly third country with no such rules, then send the goods on. Malaysia was one of the favorites. At Washington's urging it started requiring permits in July 2025 for moving advanced U.S.-origin AI chips through its territory. Singapore and Taiwan have brought in similar controls.

So the trick now is the paperwork. Call the servers something boring, pass through Malaysia as a transit point, and the restrictions at the next stop never get triggered. Tom's Hardware, reporting the seizure, called China the most likely end point. Malaysian officials have not said so.

Thirty servers is a real haul. It's also physical: crates, a declaration form, a warehouse, an airport, a customs officer who can open the box.

Meanwhile, at the API

On the same day, Reuters relayed a report from The Information that DeepSeek and Moonshot AI leaned heavily on outputs from OpenAI's GPT-6 Astra to train their newest models. Citing people with direct knowledge, The Information named Moonshot's Kimi K3.1 Pro, released last week, and an upgraded DeepSeek V4.1-Flash that shipped on October 17. None of the three companies immediately answered Reuters' requests for comment. Moonshot has previously denied building its models this way.

The technique is called distillation. You send a strong model huge numbers of prompts, collect what comes back, and use the answers as training data for your own model. Done at scale, it lets a cheaper model pick up a good share of an expensive one's skills.

OpenAI had already pointed at Moonshot once. In a September 30 post, it said people linked to the Kimi developer ran a coordinated campaign in July to pull out the hidden reasoning behind its models' answers. OpenAI said it detected related activity involving more than 15,000 users, shut it down by July 28, banned the accounts and shared what it found with other labs. Earlier this month Anthropic said new safeguards had cut the data collected from Claude in distillation attacks by around 90% to 99%.

Moonshot, for its part, says Kimi K3.1 Pro beat GPT-6 Astra on a coding benchmark. Nobody has independently checked that, and there's an obvious awkwardness in a student topping the exam it allegedly copied from.

Two leaks, one border

Put the two stories side by side and the shape of U.S. AI policy looks lopsided. Hardware gets permits, transit rules, and customs officers in three countries. Model outputs go out through the front door to anyone with an account and a payment method, and the labs police that traffic themselves, after the fact, by spotting patterns in usage logs.

The two problems do connect. Distilled data is useless without hardware to train on, and training on millions of collected answers takes a lot of compute. Which is part of why someone keeps trying to move 30 servers at a time through Kuala Lumpur with the word "electronics" on the label.

Customs can open a crate. Nobody has built the equivalent for a chat window yet.

Sources: Tom's Hardware, heise online, Reuters (via Galveston County Daily News), The Next Web.

// RELATED_SIGNALS

More from the feed