Hackers Stealing Anthropic AI Claude Tokens - Vibe Coders are Cooked

Eli the Computer Guy (0:00)

Oh, lordy, lordy, lordy, the world is about to get a hell of a lot stupider. So I do think this story is kind of interesting because we have heard about the distillation attacks being used against American AI tech companies by the Chinese AI companies. And one of the big questions that I always have is just simply, how are the Chinese companies able to access so many sessions of these American AI services, right? They talk about hundreds of millions of sessions that these Chinese companies are using in order to distill the outputs from the US AI companies. And one of the big questions out there is just how are they able to access so many sessions? And so an interesting thing, it looks like it's happening now, is just like in previous generations, you had hackers going out there and stealing things like credit cards, so that criminals were able to use your credit card in order to buy things. Apparently, at this point in time, we now have hackers doing things like hijacking your sessions or stealing your AI API keys to then use them for other purposes or sell them off, which could be an interesting thing. I do think about this with the modern world of developers, back when I was learning technology, one of the things that we were told in one of our classes was basically, good administration is good security. Good security is good administration. When you are building your systems, you need to have an eye for security, just in due course. That's just what you do. One of the things I think is interesting now with the modern world of technology professionals is just that overall systemic view of systems, the holistic view of systems seems to have simply dropped out of the education system, right? People learn their little JavaScript, or they learn their little framework, or whatever else, and they don't actually understand a lot more about technology. It is one of the frustrating things that I find, like I go to technology events and that type of thing, and I am actually surprised how pillion-hold many technology professionals are nowadays, where they literally, they know what they know. To be clear, they might be great at what they know, but they know one little sliver of technology, and that's it. And so one of the big issues now is with all of these developers going out there, they're deciding to vibe code, right? Like previously, two years ago, they would have never thought about being full stack developers, but now that they have AI to help them, they can definitely build an entire stack. But the problem again is, if you're talking about being a full stack developer, not only do you need to know languages, not only do you need to know databases, but you also need to know things like permissions and security and firewall rules, and all kinds of other stuff that goes along with actually being a full stack developer. And I think that's one of the things that gets left out of the education for a lot of the young coders out there. And so one of the problems right now is they're downloading, installing all of these AI tools. They don't really understand how security works, and hackers realize that fact. And so basically, what's happening right now is hackers are stealing the tokens from actual subscribers from these different AI services and using them for their own purposes. On August 4th, Grant D. Swart, an independent AI consultant in East Sussex, UK, noticed something strange going on with his Cloud Max20X account. He hadn't been working that day, yet his token usage was climbing. The next day, he disabled everything. He had attached to Cloud and did not work with it. Token consumption again increased, quote, in the clearest controlled interval. It increased from 45% to 55%. While I performed no work, scheduled co-work tasks were paused or completed, dispatch Cloud execution was disabled, and there was no corresponding active local Cloud code task. What was eating up his token allowance? He had no idea, so he contacted Anthropic and asked for an itemized list. Anthropic didn't provide one, but it agreed something was off. It suspended his paid account and validated all of his sessions and server-side Cloud code tokens and issued him a partial refund of 44 pounds and 49 pence for the remaining time on his $200 per month subscription. I think this is kind of an interesting thing, too, to be thinking about. Again, log files, right? You can't audit what you don't log. But one of the big questions for all of these AI frontier labs is their logging systems seem to be a bit shit.