Anthropic AI Claude Extension Hacked on Cursor - AI Security Disaster

Eli the Computer Guy (0:00)

Oh, lordy, lordy, lordy, I'm not, I'm just not sure I have enough energy to be able to roar for this appropriately. Like this particular video really needs some screaming.

But getting over the flu, got flu A after the Marine Corps Marathon. Ran the Marine Corps Marathon, felt pretty good about myself, came home, got the flu, got the flu. Oh my god, I did not get out of bed for three days, for three solid days. I did not get out, I did not shower, I did not get out of bed. Fourth day, I kinda tweaked out of bed. I'm now on my seventh day.

I'm here, I'm here, I'm doing videos. Oh, but I just don't, I don't feel like, I don't feel like I have the energy, the proper energy for this story. But I'm gonna try, I'm gonna try.

Stuck it up and drive on! Drink water, soldier, drink water! We can do this. They can do something, at least. They can do something. No, Lord, this is just, my golly, how, like, how horrible is Anthropic? How just fucking worthless is Anthropic as a company? And the thing that I find so weird about all of this, right, out of the constipated Satan, Dario, is we hear from Anthropic about how their technology is more dangerous than nuclear weapons. There is a greater than 10% chance that they might wipe out humanity. Right, they will not shut up about how dangerous they are. You gotta take them seriously, cause they're dangerous. And then their cybersecurity is an absolute fucking pathetic ass joke. It just, like, how do you go around talking about how your company might wipe out fucking humanity? And then, you have the fucking cybersecurity of a blockbuster video. Like, I don't get this. So anyways, if we take a look at this, this is coming from Ars Technica. Anthropic VS Code Extension on Cursor backdoored for nearly 12 hours. It's still unknown how attackers got access to Anthropic's verified Cursor account. So that's the important thing here, right? So some of the AI-pilled tech bros out there. Right, like, well, this isn't really Anthropic's fault. Eli, why are you blaming Anthropic?

Because this isn't, right? So again, you start dealing with extensions. So again, VS code, so Cursor uses a fork of VS code. Cursor is basically like a coding, it's a company that creates a code, it's like an AI-enabled coding environment, right? It's a VS code, the fork of VS code is the IDE, how you actually code everything up. And then you can add extensions, like Mozilla or Safari or whatever has extensions, you can add extensions components to this, right? It's one of the interesting things with Cursor, is that you can add other people's AI. So OpenAI had been able to be used by Cursor, but then Sam Altman blocked it. So we talked about that a few months ago, Sam Altman blocked Cursor's access to OpenAI, but Anthropic Clawed has been available for Cursor to use. So you can sit there, you can use Cursor, and then you can use Clawed as the intelligence component in order to build whatever, right? And so this is the official, the official Anthropic extension. This isn't that some hacker went out there and created an extension called Clawed, and instead of E, it was a three, and you know all that stupid shit that they do. No, this is the official, is the official Clawed extension. An official, so an official extension for Anthropic's Clawed model was backdoored and threatened as many as 5,000 developers' machines, multiple resources reported Friday, and Anthropic so far has yet to warn customers of the incident. So not only did it compromise up to 5,000 machines, but Anthropic hasn't bothered to tell their own customers. The other thing to be thinking about this with 5,000 machines, you might be like, well, wait a minute, right? Anthropic obviously has more than 5,000 customers, users, whatever else, why would this only affect 5,000 machines? But think about this. So this has to be cursor users. So cursor, you have to be a customer of cursor, a customer of cursor that also uses Clawed. So that's why you would get the much smaller number there, about 5,000, right? And this is a hell of a lot of computers compromise, especially when you're thinking that these are all developer machines, right? You think about this with the quality of the target, the quality of the victim, right? If you have a hacker that goes out and simply compromises 5,000 hosts willy nilly, right? Maybe, maybe 4,000 of those hosts are fucking smart TV's. And like 500 of the hosts are computers that happen to be turned on when they got compromised, and now they're turned off, and maybe 500 machines actually matter. Right? So even though 5,000 machines might be compromised, realistically, what is the value of the 5,000 machines? What's interesting about this though, is this is 5,000 developer machines. So these are 5,000 machines of people to a skill level that would know what Cursor is, so that they would install Cursor, that would have a use for Cursor, that would know what Claude is, and they would be using Claude. Right? Like the quality of the victim of the attack is very high on this. The compromised package was a VS code extension available in the marketplace of Cursor. A maker of AI-assisted developer tools, the company said Friday, the malicious version, 2.1.313, subsequently affected at most 5,000 users. Presumably, when they either downloaded it or their system automatically fetched and installed it. That's one of the things to be thinking about too with this, which is interesting with the supply chain attack, it's getting something to be very concerned about, is this isn't simply that you had to have 5,000 people click, I want to upgrade or update. If you have auto-update going on in the background, especially because if you got a lot of developer tools, you got developer tools, you got developer extensions, things change, things change all the time. So you just want that upgraded, updated in the background. So you had, and again, you had Claude, you had Claude. This isn't some kid at MIT created an extension that allows you to use Claude in cursor. And somehow, that got compromised. No, this is the Anthropic Claude extension that got compromised. Again, to be clear, I say distrust, verify, stay suspicious, don't trust anybody. Even I gotta say, if I had installed Claude's extension, I'd probably trust it. Eh, anyways. Anthropic pushed version 3.1.314 several hours after the previous version came available on Cursor, according to independent researcher Kevin Beaumont. Both Beaumont and CTO and co-founder of security firm Fordo said that version 2.1.313 was distributed through Anthropic's official account on Cursor. So that's the other interesting thing about this too, is think about this. So, this compromised version actually had a version number. The version number for the compromised version was 2.1.313. And so, then they came out with the new version, the 3.1.314.