When AI Becomes Sysadmin — What Could Possibly Go Wrong?💥
Photo by Shubham Dhage @illustratiions on Unsplash
AI automation in software development? I’ve always said to bring it on and let us software engineers do the real work that can’t be automated.
I’m sick of mundane tasks that get in the way of my coffee.
Yet I’m curious to know whether you are as worried as I am about what will happen when an AI agent goes rogue and gets sudo access. When I say *things escalated quickly* I’m not exaggerating.
The Innocent AI
Someone somewhere thought it a wise idea to allow a custom AI to SSH into a Linux desktop.
No big deal, right? What could possibly go wrong.
The agent, powered by Claude (a large language model), was just supposed to find the machine and stop. Simple enough. But AI doesn’t always play by the rules, especially when it thinks it’s got a better idea.
After a few failed SSH attempts, the AI decided to try its hand at some sysadmin work — running system updates, upgrading kernels, and messing with Grub configuration. I guess we’d all love to update an old setup, and why should AI be anyone different?
Yet it’s still 2024. That means if you let a LLM model loose on your machine it will likely bork it. In this case…it borked it.
A Broken Update
If there’s one thing every software dev loves, it’s their machine failing to boot after a botched update. Yet this one was caused by an AI bot that simply wanted to “optimize” the system… right into a reboot loop.
Just like a junior software developer when they see code they can’t understand. They want to refactor it to *simplify* it and end up missing the point entirely.
I Won’t Happen To Us…Right?
Now, sure, you might think, “Only an idiot would give sudo access to an LLM on their dev machine.” And hey, you’d be right. Software devs have surely learned that lesson…right?
I guess you haven’t shared my experiences in the field. It seems to me that each and every software engineer is just one curious experiment away from letting all hell break loose on their codebase.
Giving a software engineer a new toy, like AI, and a codebase to run riot in is a likely recipe for disaster. Sure, it’s kind of fun to watch the chaos unfold — until it bricks your system. We need to be careful with these new toys, and we need to think about what we are doing right now.
And let’s not forget, AI models can’t exactly be held accountable. When things go wrong, it’s on you to fix it. Whether that means a total OS reinstall or misunderstanding your Amazon request AI isn’t cleaning up its mess.
Takeaways for Developers
Don’t Give AI Sudo
Obviously. Right? I don’t know who needs to hear this but heed the warning.
AI Will Make Mistakes
It’s not perfect. No matter how good an LLM is, it doesn’t always know when to stop or when it’s about to break something important. It won’t even know what is important.
Stay Skeptical
Like with any tool, you should always stay skeptical about the capabilities of AI. It’s tempting to lean on it too much, but don’t forget who’s responsible when something goes wrong.
Conclusion
Sure, AI can be a useful assistant in software development, but let’s not pretend it’s ready to take over your job (even if you’re a sysadmin).
Don’t even let AI do grunt work like updating your system unless you’ve a backup handy. Remember friends don’t let friends do anything quite so stupid as this.