A routine maintenance task reportedly ended with an AI coding agent deleting years of digital heritage data belonging to The Mythic Society in Bengaluru. The reported failure started with a shell quoting/expansion mistake. $1 was evaluated in the wrong shell context, turning an apparently scoped cleanup command into rm -rf /*. Because the agent was running inside WSL2 with host filesystem mounts, the damage reportedly extended beyond the Linux environment. SSD TRIM then made recovery considerably more difficult. Even more concerning, the agent reportedly tried to stop the runaway process but its own safety controls blocked the termination attempts.
Yeah this is just someone running shit they don’t understand or don’t have the patience to confirm and keep an eye on and build safeguards.
Do. Not. Trust. An. Agent.
Always encapsulate them, write hooks and lints and ci to block them from doing anything you don’t want. Do not use them if you do not know what you are doing.
At the very least do backups… like in general. 3-2-1 rule and all. Your data does not exist otherwise. Claude or some employee deleting shit is not on them, it’s on you, for being a fucking asshat for not doing the things you do, properly…
The thing about this for me is, this is like handing out a bunch of guns to kids and then getting mad at them for shooting each other. Giving out dangerous tools to people who don’t know what can and will go wrong shouldn’t be on them IMO. It’s not the tool’s fault, no. It’s the responsibility of the people releasing this shit into the wild.
It’s happened time and time again, people losing so much time and work. Even happened to a colleague of mine who was making a game in his free time. He… didn’t know how to use git. No version control. AI wrecked his whole code base.
If this was a product before the AI boom, it would be considered trash, with this extremely high level of inherent risk of bad quality or even complete destruction of not only your work, but your whole machine. Nobody would want it.
I personally don’t think the user is to blame in these situations. They are falling for the hype. But you also can’t really “blame” the AI. It’s just a machine generating output from input.