Thought You Were Using Kimi? Think Again...

Anthropic have just published a detecting and countering misuse of AI report and what came out of it is insane. Command AI. Settle down. Get get some popcorn, boys and girls. We discovered that Moonshot AI, the company that produces the Kimmy family of models, silently forwarded custom requests to Claude instead of processing them using Kimmy. Moonshot then displayed Claude's responses to users. These users thought they were using a Kimmy model but received responses from Claude instead. And I did see this when people were asking. I don't know whether it was Kimmy to be fair, but like I do remember seeing some tweets where people were like, "I've asked what model you are and it says it's Claude." So I think people were kind of on to this to be honest. Over a 10-day period, Moonshot relayed 300,000 customer requests to Anthropic, the vast majority of which were rooted to Opus. Moonshot used a proxy service network of 5,380 fraudulent accounts, most of which appeared to be located in Singapore and Japan. In addition to serving Claudial sponsors to their customers, Moonshot also captured and saved at least a portion of these exchanges. Moonshot built a coot extraction pipeline to extract Claude's uh chain of thought transcripts from those saved relay exchanges to train its models. Moonshot also extracted chain of thought transcripts harvested through other means. Absolutely bonkers. I don't think we need to go into anything more than that. But this that's just crazy, isn't it? Deepseek serves cord instead of its own model. So Deepseek are doing it as well. Distillation campaign by Xiaomi. Gosh, they're really throwing everyone under the bus here. Sentime Miniax in the third party reseller ecosystem. The proliferation of proxy services to circumvent anthropics access restrictions has created a secondary market through which labs can purchase or otherwise acquire harvested exchanges between users and claude. Some proxy networks both provide clawed access to users in supported regions and also have exchanges in order to sell them to other labs. For example, sense times distillation pipeline excluded included transcripts of user exchanges with clawude purchased from third-party vendors. So, they've bought the exchanges. They're not they're not um rooting it and collecting those exchanges themselves. There's literally like a whole black market. That's crazy. Yeah. Jesus Christ. This is um data. It's still data. Yeah. This is um great news for uh and Alibaba and Quen are doing it as well. Wow. Look at all these bloody Nora. So, anyway, that's bonkers. Look how look how long this Can you see how long this article is? Yeah, that is one heck of a report. So yeah, Open AI are like, "Oh man, we what does anyone want our data? Why didn't they want our model?" [laughter] Yeah, that's what they are doing. And it seems like they are not really that ahead or like they are further behind the Chinese models and they are just distilling. And just one thing the numbers I don't think they are that big like 300,000 requests that sounds like a weekend for me. [laughter] Probably. Out of seriousness I don't know. 300,000 requests is not as much. I don't know. I think yeah this is part of a larger conversation on my show command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI.