Claude Code is Spying On You

Claude have been a bit sneaky the last few months. Command AI. Now, they've had some issues with the old Chinese open-source labs, uh, stealing and and and, uh, what would you even call it? Training off of Opus or Clawed Models really. So there was an investigation done where he took a look at some really interesting things and I won't bore you with it but um it's it's an interesting technique called steganography where it's it's a technique used to hide data in plain sight. So uh he noticed that there are a few checks going on. First of all, your time zone, whether it's, you know, in in these sorts of time zones over in China, but it would, these are the two things it's trying to change. It removes uh or changes, sorry, the apostrophe on today's and changes dashes to slashes. And these are the indicators that CL Anthropic can use to know that their models are being used uh by employees of these open-source um uh labs. That's crazy. Yeah. Really, really interesting. So like they're checking the base URL cuz we'll get into it but I change the base URL quite a lot and I use a lot of my GLM subscriptions, my mini map subscriptions and things like that. So first of all it's just checking that then it will check that kind of what is the base URL? The base URL is it's so it should be api.anthropic.com but you can override this and and you can and this isn't illegal. you can override it because you might have a company where you're running Claude off of Bedrock which is AWS's infrastructure. So you'll change the URL to run on Bedrock. But you can also change this to point to Miniax M3. That's how you do it because that's how you do it. Yeah, you just you just set this you set this in your settings and it will just and Claude will accept it and just use that. It's not there's nothing wrong with that. But then it goes into these additional checks to check if the time zone matches. Um it also checks uh so yeah see so it's it's checking for those Moonshot which is um GLM and all this kind of stuff. Alibaba which is Quen um but also just does a few more little checks here or there and confirmed this. This is an exp because uh someone reported on it here uh breaking anthropic has been hidden spyware. This is an experiment we launched in March was meant to prevent account abuse and authorized unauthorized resellers and protect against distillation. So he's confirmed it. They were absolutely doing this. Now the reason resellers is there like a I I heard like there is a black market for tokens. So yeah, like can you buy clawed subscription for cheaper? How does that work? Well, it would just be a case of giving people your API key, right? Giving them access to an API key and then they've got access because the thing is uh some countries who is paying for that? Well, I mean imagine again it's like you might find I might get the $200 a month plan but might actually be okay for two junior developers. So it's like I'll give them an API key. So I don't think there's nothing that's probably against terms of service but I don't know but it just all maybe it's just better better for them to understand the landscape of things. So um yeah um the issue being is that uh the the fact of the matter is that they're putting tracking uh telemetry and things like that in the API because obviously before it reaches the model it goes to anthropics API and that's where a lot of the data collection is and a lot of the quality of life things like memory and stuff like that are collected in that middle API And when we're sending off, you know, I'm sure you're you you do it. You'll give it an API key to put in the EMV file or something like that. Like I try not to do that, but we are and a lot of people are chatting personal information and things like that that you don't you want to feel you want to know you want to know what Anthropic are tracking. If they're going to track, you want to know what they're tracking. And I don't think this was revealed, you know, I don't think this they were honest about this and yeah, you know that that this is this is um what's the was it Cambridge Analytica not to the same extent but it's like you know you say one thing but you're actually doing another. I think that's where a lot of the annoyances come from. Uh so just a short note on reading your environment file N file like doesn't do this by default right can it do it and tell you I didn't do it yeah no it it should it should respect by default if you go into claw code it should respect your git ignore file so you're having there node modules you're having there your env files right so that would respect if you don't have a git ignore more then it there's probably some training data to say don't look in it but it I've had it look in like EMV file. Yeah, I I remember like I had literally last night it was telling me uh I'm not going to look at it. I'm creating a duplicate of the env file and I'll show you like where to go and where to put your stuff. And then I had some difficulties and it was like I'm basically the model being done with me like I just did it for you. Yeah. But like you needed to give it the API key for it to put it in there, right? Yeah. Yeah. Yeah. So it's by that point it's passed, you know, it's it's in the model. But yeah, so yeah, flip of a coin. I've had it ignore it. I've had it look into it. you know, it looked into it last night cuz I was like trying to figure out some stuff. Um, I don't think I even prompted it to look at it and it did. So, it's definitely it's definitely going going through their thing. Um, but 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.