Is this really that surprising? It used to, still is?, be considered a bad idea to follow those install suggestion where it’s `/bin/bash -c $(curl example.com/install.sh)`. These tools are basically that but even more risky because the code they generate is semi-random.
Of course you can evaluate everything it does before it does it. But no one reads the homebrew install script to make sure it’s safe when setting up a new Mac so who’s going to read the Claude code scripts?
That's what I don't get about the current LLM coding/agent hype. The supposed gain is completely negated unless you can trust the output completely. If I have to inspect everything, generating a bunch of code/scripts or doing a lot of things/running a lot of commands quickly on my computer doesn't really help at all. The more things it does the more of a reasoning burden I have.
What I think is completely missed is that the bottleneck in knowledge work is understanding and making sure about what is built and done.
It's a bit funny that between trolling and innocent "screw it, burn everything and start over" that every beginner has been through, this is obviously going part of the learning experience... whether you're human or machine.
(No, I wouldn't seriously rm -rf ~/ intentionally in most cases but it's not so insane when you imagine creating a new user for project separation. For those reasons it's probably good to keep the agents separate too...)
Claude wrote a Python file earlier to delete/clean up some folders. It was about to run, so I canceled the run. It then tried to write a bash script to delete the files. I canceled it and then tried to write the command to delete them. I was amazed by all the ways it was finding.
I didn’t read the word “directory” initially, and I was imagining a scenario where the author gave Claude access to an API, and was like “this endpoint will set off a nuke” and then Claude was like “Ok! :3” and triggered it.
I've had situations where Cursor just starts to do some really bizarre behavior after long running cycles of tasks unsuccessfully like the death loop I've seen described in other threads.
Best way to deal with this is to just clear the embedding index from the cursor settings and rebuild it.
I've never had it go to a point where it will want to rf home, but now I'm a bit fearful that one day it will go and do it as I have it on auto run currently.
Really helps avoid these kinds of situations. I do the same thing with `cp`, but unlike `rm`, `--force` doesn't actually undo `--interactive` in `cp` so I have to call `cp` directly (e.g., `/usr/bin/cp`) when I don't want to deal with the prompts.
(On Mac and other systems with BSD versions of tools, `--interactive` is `-i` and `--force` is `-f`.)
'--force' does more than undoing '--interactive'. '--force' will remove a file even if the permissions of that file are set to read only (if you own the file or directory):
> touch foo
> ls -l
-rw-r--r--. 1 foo foo 0 Mar 24 05:54 foo
> chmod a-w foo
> ls -l
-r--r--r--. 1 foo foo 0 Mar 24 05:54 foo
> rm foo
rm: remove write-protected regular empty file 'foo'? n
> rm -i foo
rm: remove write-protected regular empty file 'foo'? n
> rm -i -f foo
>
my consequence is that i never alias 'rm' to default to '-i', because 'rm -f' is dangerous. instead i type 'rm -i' manually every time, and remove the '-i' if i don't want it.
For context, according to the rules of the subreddit you must provide the full sequence of prompts used to generate the output, but no such sequence is given.
Also Claude didn't "try" to do anything of course. Such a title is nonsense at best.
Readit News
Of course you can evaluate everything it does before it does it. But no one reads the homebrew install script to make sure it’s safe when setting up a new Mac so who’s going to read the Claude code scripts?
What I think is completely missed is that the bottleneck in knowledge work is understanding and making sure about what is built and done.
(No, I wouldn't seriously rm -rf ~/ intentionally in most cases but it's not so insane when you imagine creating a new user for project separation. For those reasons it's probably good to keep the agents separate too...)
Best way to deal with this is to just clear the embedding index from the cursor settings and rebuild it.
I've never had it go to a point where it will want to rf home, but now I'm a bit fearful that one day it will go and do it as I have it on auto run currently.
Accidents will happen, disaster recovery is needed notwithstanding misaligned AI.
IMO once you're really ready for the worst, productivity from yolo mode AI assistance surpasses the downsides.
(On Mac and other systems with BSD versions of tools, `--interactive` is `-i` and `--force` is `-f`.)
Deleted Comment
Also Claude didn't "try" to do anything of course. Such a title is nonsense at best.
It just continues to do things... That's the value prop.