In my opinion this is a solution at the wrong layer. It's working by trying to filter executed commands, but it doesn't work in many cases (even in 'strict mode'), and there's better, more complete, solutions.
What do I mean by "it doesn't work"? Well, claude code is really good at executing things in unusual ways when it needs to, and this is trying to parse shell to catch them.
When claude code has trouble running a bash command, it sometimes will say something like "The current environment is wonky, let's put it in a file and run that", and then use the edit tool to create 'tmp.sh' and then 'bash tmp.sh'. Which this plugin would allow, but would obviously let claude run anything.
I've also had claude reach for awk '{system(...)}', which this plugin doesn't prevent, among some others. A blacklist of "unix commands which can execute arbitrary code" is doomed to failure because there's just so many ways out there to do so.
Preventing destructive operations, like `rm -rf ~/`, is much more easily handled by running the agent in a container with only the code mounted into it, and then frequently committing changes and pushing them out of the container so that the agent can't delete its work history either.
Half-measures, like trying to parse shell commands and flags, is just going to lead to the agent hitting a wall and looping into doing weird things (leading to it being more likely to really screw things up), as opposed to something like containers or VMs which are easy to use and actually work.
I recently had a similar conflict with GPT-5.1, where I did not want it to use a specific Python function. As a result, it wrote several sandbox escape exploits, for example the following, which uses the stack frame of an exception to call arbitrary functions:
Yeah, I had an issue where Claude was convinced that a sqlite database was corrupt and kept wanting to delete it. It wasn't corrupt, the code using it was just failing to parse the data it was retrieving from it correctly.
I kept telling it to debug the problem, and that I had confirmed that database file was not the problem. It kept trying to rm the file after it noticed the code would recreate it (although with no data, just an empty db). I thought we got past this debate until I wasn't paying enough attention and it added an "rm db.sqlite" line into the Makefile and ran it, since I gave it permission to run "make" and didn't even consider it would edit the Makefile to get around my instructions.
If the LLM never gets a chance to try to work around the block then this is more likely to work.
Probably one better way to do this would be, if it detects a destructive edit, block it and switch Claude out of any autoaccept mode until the user re-engages it. If the model mostly doesn't realize there is a filter at all until it's blocked, it won't know to work around it until it's kicked the issue up to the user, who can prevent that and give it some strongly worded feedback. Just don't give it second and third tries to execute the destructive operation.
Not as good as giving it a checkpointed container to trash at its leisure though obviously.
I agree with this take. Esp with the simplicity of /sandbox
I created the feature request for hooks so I could build an integrated governance capability.
I don’t quite yet think the real use cases for hooks has materialized. Through a couple more maturity phases it will. Even though it might seem paradoxical with “the models will just get better” - to which is exactly why we have to be hooked into the mech suits as they'll end up doing more involved things.
But I do pitch my initial , primitive, solution as “an early warning system” at best when used for security , but more so an actual way (opa/rego) to institute your own policies:
I got hooks working pretty well for simpler things, a very common hello world use case for hooks is gitleaks on every edit. One of the use cases I worked on for quite awhile was getting hooks that ran all unit tests at the end before the agent could stop generating. This approach forces the LLM to then fix any unit tests it broke and I also enforce 80% unit test coverage in same commit. I found it took a bit of finagling to get the hook to render results in a way that was actionable for the LLM because if you block it but it doesn’t know what to do it will basically endlessly loop or try random things to escape
FWIW I think your approach is great, I had definitely thought about leveraging OPA in a mature way, I think this kind of thing is very appealing for platform engineers looking to scale AI codegen in enterprises
Part of my initial pitch was to automate linting. Interesting insight on the stop loop. Ive been wanting to explore that more. I think there is a lot to be gained also with llm-as-a-judge hooks (they do enable this today via `prompt` hooks).
I dont think the team meant for the hooks to work with plan mode this way (its not fully complete with approve/allow payload), but it enabled me to build an interactive UX I really wanted.
I think the key you point out is something that is worth observing more generically - if the LLM hits a wall it’s first inkling is not to step back and understand why the wall exists and then change course, its first inkling is to continue assisting the user on its task by any means possible and so it’s going to instead try to defeat it in any way possible. I see the is all the time when it hits code coverage constraints, it would much rather just lower thresholds than actually add more coverage.
I experimented with hooks a lot over the summer, these kind of deterministic hooks that run before commit, after tool call, after edit, etc and I found they are much more effective if you are (unsurprisingly) able to craft and deliver a concise, helpful error message to the agent on the hook failure feedback. Even just giving it a good howToFix string in the error return isn’t enough, if you flood the response with too many of those at once the agent will view the task as insurmountable and start seeking workarounds instead.
> ... if the LLM hits a wall it’s first inkling is not to step back and understand why the wall exists and then change course, its first inkling is ...
LLM's do not "understand why." They do not have an "inkling."
Claiming they do is anthropomorphizing a statistical token (text) document generator algorithm.
The more concerning algorithms at play are how they are post-trained. And the then concern of reward hacking. Which is what he was getting at.
https://en.wikipedia.org/wiki/Reward_hacking
100% - we really shouldn't anthropomorphize. But the current models are capable of being trained in a way to steer agentic behavior from reasoned token generation.
> But the current models are capable of being trained in a way to steer agentic behavior from reasoned token generation.
This does not appear to be sufficient in the current state, as described in the project's README.md:
Why This Exists
We learned the hard way that instructions aren't enough to
keep AI agents in check. After Claude Code silently wiped
out hours of progress with a single rm -rf ~/ or git
checkout --, it became evident that "soft" rules in an
CLAUDE.md or AGENTS.md file cannot replace hard technical
constraints. The current approach is to use a dedicated
hook to programmatically prevent agents from running
destructive commands.
Perhaps one day this category of plugin will not be needed. Until then, I would be hard-pressed to employ an LLM-based product having destructive filesystem capabilities based solely on the hope of them "being trained in a way to steer agentic behavior from reasoned token generation."
Just put it in a container. I use bash aliases like this to start a throwaway container with bind mounted cwd, works like a charm with rootless podman. I also learned to run npm and other shady tools in this way and stopped worrying about supply chain attacks.
alias dr='docker run --rm -it -v "$PWD:$PWD" -w "$PWD"'
alias dr-claude='dr -v ~/.claude:/root/.claude -v ~/.claude.json:/root/.claude.json claude'
I do that, too! I use git for version control outside the docker container, and to prevent claude from executing arbitrary code through commit hooks, I attach the docker volume mount in a nested directory of the repository so claude can not touch .git. Are there any other attack vectors that I should watch out for?
Ohh, good point about git hooks as a container escape vector! I probably should add `-v $PWD/.git:$PWD/.git:ro` for that (bind-mount .git as read-only).
I always run my agents in a container with the source code directory mounted. That way I can reasonably be confident I may let it work without fearing destructive actions to my system. And I'm a git reset away to restore source code.
Someone should write a version of this that uses AI to detect whether the command that the AI wants to run is dangerous. Certainly that seems like the current trend in software "engineering".
Right? The training set must be insane. The way it heads/tails/greps to limit tokens ingested must have taken a lot to train — that's not something one finds on SO
Two MCP tools back to back on the HN frontpage when seemingly dozens of them doing the same functionality already exist. Both posts written by AI with the typical tells. Daring today aren't we?
You should probably rely less on AI. If your first thought is "I need to delete some directories" and your immediate next thought is "I'd better ask an AI agent to do this for me", you are definitely exhibiting skill entropy.
Claude does these things even though you have explicit instructions not to do them, this isn't a tool for you asking it to delete files.
Just today Claude decided to do a git restore on me, blowing away local changes, despite having strict instructions to do nothing with git except to use it to look at history and branches.
Why jump to the conclusion that the person is so incompetent with no evidence?
Because there's now a class of programmers who are very anti AI when it comes to coding because they think anybody who relies on it are degenerate vibe coders who have no idea what they are doing. You can see this in pretty much every single HN post w.r.t AI and coding.
Skill entropy is a result of reliance on tools to perform tasks which otherwise would contribute to and/or reinforce a person's ability to master same. Without exercising one's acquired learning, skills can quickly fade.
For example, an argument can be made that spellcheckers commonly available in programs degrade people's ability to spell correctly without this assistance (such as when using pen and paper).
Thanks for framing my physical disability as a skill issue. Injuries i sustained developing my skills beyond what most others were willing to do, but i guess my use of AI to assist my input so i can continue developing totally erases that experience.
What do I mean by "it doesn't work"? Well, claude code is really good at executing things in unusual ways when it needs to, and this is trying to parse shell to catch them.
When claude code has trouble running a bash command, it sometimes will say something like "The current environment is wonky, let's put it in a file and run that", and then use the edit tool to create 'tmp.sh' and then 'bash tmp.sh'. Which this plugin would allow, but would obviously let claude run anything.
I've also had claude reach for awk '{system(...)}', which this plugin doesn't prevent, among some others. A blacklist of "unix commands which can execute arbitrary code" is doomed to failure because there's just so many ways out there to do so.
Preventing destructive operations, like `rm -rf ~/`, is much more easily handled by running the agent in a container with only the code mounted into it, and then frequently committing changes and pushing them out of the container so that the agent can't delete its work history either.
Half-measures, like trying to parse shell commands and flags, is just going to lead to the agent hitting a wall and looping into doing weird things (leading to it being more likely to really screw things up), as opposed to something like containers or VMs which are easy to use and actually work.
I kept telling it to debug the problem, and that I had confirmed that database file was not the problem. It kept trying to rm the file after it noticed the code would recreate it (although with no data, just an empty db). I thought we got past this debate until I wasn't paying enough attention and it added an "rm db.sqlite" line into the Makefile and ran it, since I gave it permission to run "make" and didn't even consider it would edit the Makefile to get around my instructions.
Probably one better way to do this would be, if it detects a destructive edit, block it and switch Claude out of any autoaccept mode until the user re-engages it. If the model mostly doesn't realize there is a filter at all until it's blocked, it won't know to work around it until it's kicked the issue up to the user, who can prevent that and give it some strongly worded feedback. Just don't give it second and third tries to execute the destructive operation.
Not as good as giving it a checkpointed container to trash at its leisure though obviously.
I created the feature request for hooks so I could build an integrated governance capability.
I don’t quite yet think the real use cases for hooks has materialized. Through a couple more maturity phases it will. Even though it might seem paradoxical with “the models will just get better” - to which is exactly why we have to be hooked into the mech suits as they'll end up doing more involved things.
But I do pitch my initial , primitive, solution as “an early warning system” at best when used for security , but more so an actual way (opa/rego) to institute your own policies:
https://github.com/eqtylab/cupcake
https://cupcake.eqtylab.io/security-disclaimer/
FWIW I think your approach is great, I had definitely thought about leveraging OPA in a mature way, I think this kind of thing is very appealing for platform engineers looking to scale AI codegen in enterprises
Ive had a lot of fun with random/creative hooks use cases: https://github.com/backnotprop/plannotator
I dont think the team meant for the hooks to work with plan mode this way (its not fully complete with approve/allow payload), but it enabled me to build an interactive UX I really wanted.
It will just as easily get around it by running it as a bash command or any number of ways.
I experimented with hooks a lot over the summer, these kind of deterministic hooks that run before commit, after tool call, after edit, etc and I found they are much more effective if you are (unsurprisingly) able to craft and deliver a concise, helpful error message to the agent on the hook failure feedback. Even just giving it a good howToFix string in the error return isn’t enough, if you flood the response with too many of those at once the agent will view the task as insurmountable and start seeking workarounds instead.
LLM's do not "understand why." They do not have an "inkling."
Claiming they do is anthropomorphizing a statistical token (text) document generator algorithm.
100% - we really shouldn't anthropomorphize. But the current models are capable of being trained in a way to steer agentic behavior from reasoned token generation.
This does not appear to be sufficient in the current state, as described in the project's README.md:
Perhaps one day this category of plugin will not be needed. Until then, I would be hard-pressed to employ an LLM-based product having destructive filesystem capabilities based solely on the hope of them "being trained in a way to steer agentic behavior from reasoned token generation."Just containerize Claude.
How is this not common practice already?
Are people really ok with a third party agent running out of their home directory executing arbitrary commands on their behalf?
Pure insanity.
The problem seems to come when it’s stuck in a debug death loop with full permissions.
Just today Claude decided to do a git restore on me, blowing away local changes, despite having strict instructions to do nothing with git except to use it to look at history and branches.
Why jump to the conclusion that the person is so incompetent with no evidence?
https://www.mdpi.com/2075-4698/15/1/6
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4812513
Skill entropy is a result of reliance on tools to perform tasks which otherwise would contribute to and/or reinforce a person's ability to master same. Without exercising one's acquired learning, skills can quickly fade.
For example, an argument can be made that spellcheckers commonly available in programs degrade people's ability to spell correctly without this assistance (such as when using pen and paper).