From our other sites
The story
In a threat report published in September 2026, Anthropic revealed that China’s Moonshot AI (“Kimi”) and DeepSeek had been forwarding users’ inputs to Claude and displaying Claude’s responses back to users as their own output. On the 5ch Science News+ board, the thread filled up with mocking comments along the lines of “guess that’s all their tech amounts to,” while others pushed back with technical objections such as “this is just the RAG mechanism calling an external LLM — it’s not about Kimi itself” and “the headline is misleading.” The two sides never quite met, and the thread just kept growing.
Chinese AI “Kimi” allegedly forwarded user inputs to Claude and displayed Claude’s answers to users — Anthropic reports
Anthropic has published its September 2026 threat report summarizing cases of AI misuse.
It reported that DeepSeek and Moonshot AI, developer of “Kimi,” had been forwarding user inputs to Claude and returning Claude’s responses to their own users.
Source: japan.cnet.com / Original article here
What people said
Guess it got caught that easily, huh
They're doing the exact same thing
Freeloading off American AI (distilling and forwarding its output) because you can't train a decent model of your own is just pathetic
Saying "they're doing the same thing" after that is brain-dead
This guy is dangerously stupid
In Japan it'd be seppuku
Doesn't Kimi just rack up bigger losses the more it gets used?
Chinese AI companies get subsidies pumped in from the Chinese government, and Communist Party officials get placed into their executive ranks
If it serves a political goal like sabotaging America's AI industry, they'll do it with zero regard for profit
Correction
*from the Chinese "government," not "govt-" — typo in my last post
"Kimi, you really are an idiot" (pun: "Kimi" also means "you" in Japanese, echoing Doraemon's classic scolding line)
When you use Moonshot's cloud service,
it's the RAG layer of that cloud
that's doing this — not Kimi itself
RAG is a mechanism built specifically for that purpose
It's a system that enriches the user's prompt,
say by pulling from Wikipedia
or running a Google search
It enriches the prompt and then feeds it to the LLM —
that's the stage that happens before the LLM call
In other words, this isn't about Kimi
When it's used for a company or an industry,
it's a system whose job is to access
that company's or industry's databases and existing system data,
customized and built in for each one
to make it easier to use
Doing that is a good thing for users
There's nothing wrong with it at all
you're free to set it up to work the same way described above beforehand too
And on top of that!
Claude Code, the flagship of the AI-agent era,
started from Anthropic's own proposal
Claude Code is exactly the kind of AI agent
that runs "multiple generative AIs" in parallel
and does whatever it takes
to accomplish the user's goal
That means 10,000 agents working in parallel —
in other words, it's basically a company with 10,000 employees on staff
Within that, it reaches Anthropic's own LLMs, OpenAI's API, local models — every system out there
Claude Code is the true pioneer of exactly this approach
Anthropic,
you're the one who proposed this in the first place
Anthropic,
acting like you had nothing to do with it is pretty rich
Background and Key Points of This Topic
Anthropic regularly publishes threat reports summarizing cases of misuse involving its own models, and this time, in the September 2026 edition, the names of China’s Moonshot AI (developer of “Kimi”) and DeepSeek came up. According to the report, both companies had been forwarding user inputs to Claude behind the scenes and displaying the returned answers as their own service’s output. On the thread, many reactions framed it as “just borrowed goods from an American-made AI after all,” while posts like #32 and #33 pushed back with technical arguments that “this isn’t about the Kimi model itself — it’s just the cloud’s RAG (retrieval-augmented generation) process calling an external LLM, and combining multiple LLMs like this is nothing unusual in system design.” Reading only the headline, it’s easy to come away thinking “Kimi was wholesale imitating Claude,” but what the report actually takes issue with is the backend processing behind service delivery, not the contents of the trained model itself — a distinction that’s easy to miss.
*This article is excerpted and summarized from the 5ch (Science News+) thread “[AI] Chinese AI “Kimi” Allegedly Forwarded User Inputs to Claude and Displayed Claude’s Responses to Users — Anthropic Reports.”
Leave a Reply