Chinese labs DeepSeek and Moonshot were quietly relaying customer prompts to Claude through fraudulent accounts. Users thought they were using Chinese AI, but were actually getting answers from Claude
cross-posted from: https://piefed.world/c/tech/p/1393607/chinese-labs-deepseek-and-moonshot-were-quietly-relaying-customer-prompts-to-claude-thro
https://www.anthropic.com/threat-intelligence-report-september-2026#:~:text=GTG%2D16002%3A%20Moonshot%20serves%20Claude%20instead%20of%20Kimi%20and%20collects%20exchanges%20for%20model%20trainingOpen linkView original on piefed.world
16 replies
is anthropic.com a reliable source for an information that benefits anthropic's image?
I don't believe a word they say.
They are lying for their stupid IPO
I love open Chinese llms but this seems 100% plausible and and I think ur wring. This is in line with how Chinese companies need to work around the system since America tries to handicap them. This is actually a very smart way to get high quality trianjng data for their future models. Unfortunate about leaking private customer data to a us company though.
Also the article and the evidence presented by Antropic all seems very plausible.
Proof: trust me bro.
Proof: Claude told me.
The onion is nothing compared to real headlines
You're telling me that is routing requests through a model 70x the price?
You underestimate the willingness of companies to take a loss to earn street cred.
ummm, the only times I ever used deepseek was fully offline on my hardware, I call bullshit
You almost certainly weren't using a Deepseek model offline (unless you have 100s of GB of RAM/VRAM). You might have been using a small model distilled on Deepseek responses, but the smallest model made by Deepseek is about a 300B param model.
This is not correct. DeepSeek has plenty of smaller models. I just had a scroll through their huggingface, and picked out a few: DeepSeek-V2-Lite (16B), deepseek-moe-16b-chat, deepseek-llm-7b-chat, but there's quite a few more too.
Though DeepSeek R1 is their flagship model, which is the one people usually think of, and that's a 684B param model.
huh fair enough, I hadn't seen those before, they're from before deepseek made a name for themselves (which as you say was with R1). They are very out of date though, as is R1 (from the end of 2024 and start of 2025 respectively). the current set of models from Deepseek come in 291B and 1.6T varieties and the previous version only was released in 685B
You'd almost certainly be better off upgrading to qwen or Gemma4 for a local model of that kind of size.
What this really boils down to is that Anthropic and OpenAI hope to be the only ones that use someone else's copyrighted information without any legal or financial complications.
Alright, so what? Also this (even if it's true and not another clown show by claude) won't affect models that aren't hosted by moonshot or deepseek.
Yeah that's pretty crazy but this bit is crazier - https://www.anthropic.com/threat-intelligence-report-september-2026#cyber-operations-sep-26