Transcription
Today is a really exciting day. We are onboarding our first AI employee. His name is Jerry. He's living on this brand new Mac Mini and he is powered by Open Claw, formerly Cloudbot. So, we're going to go through a complete onboarding from a security perspective. And those of you who have set up enterprise accounts for other employees, this is going to look very similar to that.
And I wanted to really take some time to talk about this today because this list has been floating around the internet of all of the critical issues with open claw or clawbot. Right? We have SSS uh brute force issues, exposed control gateways, we've got Telegram allow list, do not allow list vulnerabilities, browser hijacking, we've got password manager extractions, Slack token theft. Um, we've got root access without sandboxing. Uh, we've got prompt injections. We've got malicious skills. And, you know, any combination of all of these can really take what could be an incredibly powerful assistant and um, make you wish you never did it to begin with. So, let's get started.
Also want to put out a disclaimer. This is not a endall direction for a secure install. Um, this is still really new stuff. We are at the bleeding edge of this today and in particular as more extensions come out, more capabilities come out. You really have to take those each with its own grain of salt and understand how that fits into the larger security posture. So, buyer beware on this. Um, but this is a really good spot to start if you want to begin experimenting. All right, let's dig in.
Mac Mini box opening. All right, who doesn't love opening a brand new piece of hardware? I would be lying if I said this wasn't a big motivation cuz we get to play with a brand new Mac Mini. So, here's what the setup will look like. The Mac Mini will be on the server and Jerry will be installed inside a virtual machine on the server. This is principle one, VM isolation. If anything goes wrong, Jerry will be confined to a virtual machine and he won't get kernel access to the server.
Principle two is network segmentation. We'll put a firewall on the server side and we'll configure it to whitelist only a select few sites we need Jerry to visit to do his day job. Principle three is least privilege. Jerry will be a standard user account. No pseudo access, no separate browser profile, no access to passwords or sensitive files.
Principle four is authentication and authorization. Jerry gets his own accounts. There's no anonymous access, no shared credentials. We're going to hook him up to his own clawed or open AI login, his own email account, his own GitHub account, just like we would a human employee.
Principle five is potentially the most important is monitoring and then auditability. So if anything does go wrong, we need to be able to observe it. So we're going to log everything that Jerry does, set up alerts on anomalies, and then review this activity regularly.
So we're going to set up UTM for our virtual machine environment. And I'm going to go ahead and select Mac OS as our operating system. That's going to set up by itself. And while it's doing that, we're going to load up Lulu as our network firewall. Within Lulu, we can decide which sites we want to block, what traffic we want to allow, and it's also going to give us logging on the server side.
All right, we are ready to install Mac OSS in a virtual machine for Jerry. This is Jerry's new home. So, we're going to go through these setup steps here and making sure to lock this down by default. All right, this is going to be Jerry. We're going to give Jerry a password. We are not going to set up an Apple account. No location services. We are not going to share anything with Apple. No screen time. No Siri. Yes. Turn on file vault. We're going to write down this recovery key.
All right, let's install Jerry. Browser. Open claw.ai. Copy this. Pseudo. Paste that. Enter our password. Here we go. We're installing Jerry.
So, I quickly ran through the command line setup. I entered our API keys for Claude, for um ChattBT. We got to the UI. Right now, the UI, we can continue setting up additional features and additional connections. Um, but the reality of this is the secure setup really lives in the environment that you put this in, right? So a virtual machine that's isolated within its own server and that has endtoend monitoring logging firewall capabilities and we're giving at least privilege right we're next step on here is to remove the admin access from this particular user account before we start using and we've created our own unique clawed loginins our own unique open AI plugins that are specifically attached to Jerry. So now we can track Jerry and we can turn Jerry off if Jerry's not doing something correctly.
The reality of this is, you know, this open AI, it's still highly probabilistic and it can go and hallucinate. It can mess up the things that you're connecting it to. So if you take these steps and you get your virtual machine set up, this thing can be an incredibly powerful way for you to explore. And until somebody comes out with a truly hardened enterprise version, um I would be very cautious with how you use this. But this setup should give you a head start in at the very least minimizing the destruction if and when that does occur.
All right. Happy building.