Hacker News Reader

Comment by eterm | Hacker News Reader

This is pretty relevant for things like claude-code, which has a fairly rudimentary way of dealing with permissions with block-lists and allow-lists.

I once accidentally gave my claude "powershell" permissions in one session, and after that any time it found it was blocked from using a tool, e.g. git, it would write a powershell script that did the same thing and execute the script to work around the blocked permission.

Obviously no sane system would have "powershell" in a generic allow-list, but you could imagine some discrepancy in allowed levels between tools which can be worked around with the techniques on this page.

View on HN

Power Shell or Python scripts to work around restrictions are the go to for LLMs.

And it doesn't stop there.

Yesterday I was trying to figure out some icons issue in KDE plasma (I know nothing about KDE). Both Claude and Codex would run complex bus and debug queries and write and execute QML scripts with more and more tools thrown into the mix.

There's no way to properly block them with just allow- and block lists

> There's no way to properly block them with just allow- and block lists

Especially not when some harnesses rely on the reliability of the LLM to determine what's allowed or not, pretty much "You shouldn't do thing X" and then asking the LLM to itself evaluate if it should be able to do it or not when it comes up. Bananas.

Only right and productive way to run an agent on your computer is by isolating it properly somehow then running it with "--sandbox danger-full-access --dangerously-bypass-approvals-and-sandbox" or whatever, I myself use docker containers, but there are lots of solutions out there.

You have to be extremely careful when you set up a dev container, lock down file access, do not give the agent the power to start other containers or "docker compose up", restrict network access to an allow-list etc. Just running the agent in a container does little to protect you. (Maybe you know this, but a lot of people don't!)

Most of those things are what happens by default. Sure, be careful, but by default it's secure enough to prevent most potential issues. No need to lock down file access for example, by default it only has access to files inside the container, and of course by default containers don't have access to start other containers, and so on.

Good word of caution though, make sure you actually isolate when you set out to isolate something :)

I've just discovered and started using smolmachines^1 which actually have the requisite isolation.

1. https://smolmachines.com

As mentioned, "podman/docker run -it $my-image codex" also actually has the requisite isolation by default, no need for special software. Biggest risk is accidental deletion of stuff, easily solved without running an entire VM, which "smol" machines seems to be. No doubt VMs have their uses too, but for simple isolation like this I personally rather use already existing tooling.

Ok, YMMV, but a smolvm provides macOS-native, per-workload isolation -- vs trad container depending on a daemon and relying on namespaces (w/ a shared kernel). Easy "packing" into single-file executables, and a nice SDK, make it ~ideal for my needs; great balance of security:convenience.

https://smolmachines.com/#comparison

Cool ad bro, but stop claiming container won't get you "per workload isolation" just because they share kernels, in the context of this discussion it hardly matters, containers isolates enough for this.

ad? I have no affiliation w smolmachines, just glad I found it.

In a previous employer, they block the chmod command. I took the habit to python -c "import os; os.chmod('my_file',744)".

Glad to see LLM re-discover this trick.

> to see LLM re-discover

I imagine someone probably wrote very specifically about it in the training data that underwent lossy compression, and the LLM is decompressing that how-to.

So I'd say it's more like "surfacing" or "retrieving" than "re-discovering".

They scraped everything on Stackoverflow, likely IRC logs from Freenode, and every book written in the modern era courtesy of Sci-Hub / Library Genesis / Anna's Archive / Z Library.

RIP Aaron Swartz, they're generating trillions in shareholder value from the spiritual successors to the work they were going to imprison you for.

Indeed, I check and the solution was already on stack overflow https://askubuntu.com/a/1483248

For the LLM it's a probabilistic set of strings that achieves the outcome, the highest probability set didn't work, try the next one until success or threshold met. A human sees the implicit difference between the obvious thing not working indicating someone doesn't want you to do it, but an LLM unless guided doesn't seen that sub-text.

So chmod +x file didn't work, now try python -c "import os; os.chmod('file',744)"

Humans and LLMs both only see that when given the right context. A tool not working in a corporate environment may be anything from oversight, malfunction all the way to security block. Knowing which one it is takes a lot of implicit knowledge. Most people fail to provide this level of context to their LLMs and then wonder why they act so generic. But they are trained to act in the most generic way unless given context that would deviate from it.

Reply on news.ycombinator.com