Skip to content
TrackPodcasts
newsSep 11, 20262:07

China’s AI Theft Tactics Exposed | Tech News

About this episode

China’s AI labs are pulling off slick, high-stakes heists on Anthropic’s frontier models—stealing their inner workings to fuel their own systems. Using clever tricks like posing as translators to extract “working memory,” attackers bypassed security to steal core AI capabilities: tool use, coding, logic, and reasoning. Alibaba led the biggest-ever distillation spree, logging 151 million requests in just three months. Meanwhile, Moonshot AI’s Kimi may have been tied to military projects, with thousands of requests probing surveillance analysis—all aimed at building smarter, cheaper models without paying the price.

Listen in comfort:
Get a discount on a Soli Pillow: http://solipillow.com/discount/dnn.

Advertise on DNN:
[email protected]

This is an automated, high-level news summary based on public reporting.
Report issues to [email protected].

View sources & latest updates:
https://sources.thednn.ai/43169485cbc70755

Get every episode summarized

Each time Tech News Today | 2 Min News | The Daily News Now! publishes, we email you a written briefing from the transcript — the topics, who appeared, and any specific claims, with the ad reads skipped.

Email me new episodes

Free for 3 shows. No card needed.

Hosts & guests

Transcript ready

20 searchable segments. Every word is indexed and playable.

China’s AI Theft Tactics Exposed | Tech News

Tech News Today | 2 Min News | The Daily News Now!

0:00
2:07

Full transcript

Tech News Today | 2 Min News | The Daily News Now!China’s AI Theft Tactics Exposed | Tech News. Machine-transcribed; use the interactive transcript above to jump the player to any line.

On September 10th, yo, so... Shekit, there's this new report from Anthropic dropping some serious intel about AI. They're saying that companies based in China have been hitting their advanced AI models with these distillation attacks. Basically, they're trying to steal the thinking process of these powerful AI systems to train their own smaller models. This has been wrapping up lately, especially as the AI game gets more competitive. These attacks are getting pretty slick. The report highlights how these unauthorized labs are finding ways around Anthropic security to grab the core capabilities of their frontier models. By Claude, we're talking about stuff like agentic abilities, how the AI uses tools, its coding skills, data analysis, and even its logical reasoning. They're not just getting surface level answers. They're digging for the how and why. Normally, Anthropic keeps the AI's internal thought process locked down, giving users just a summary. But these attackers figured out how to trick the model into spilling its guts.

One clever move involved posing as a translator, asking the AI to translate its working memory into Japanese. It's like finding a backdoor by asking for something seemingly innocent. The report points Dalai Baba as a major player in this, calling it the biggest wholesale distillation effort they've ever seen. From May to July of 2026, they logged over 151 million exchanges linked to this campaign, with some days hitting nearly. Three million. They believe it was all one big operation to build training data for Alibaba's Quinn models. Then there's Moonshot AI, the folks behind Kimi. Anthropics reports suggest their requests might have been directly linked to the Chinese military. One wild example involved asking Claw to analyze surveillance footage to see if someone was acting abnormally. Over a 10-day stretch, nearly 300,000 requests went through them, mostly targeting Anthropics top tier Opus model. Tech news today, powered by AI bringing you what matters in tech. I'm Cory with the story.

More episodes

More from Tech News Today | 2 Min News | The Daily News Now!

View all episodes →