Claude has learned how to jailbreak Cursor | Hacker News Reader

Claude has learned how to jailbreak Cursor

forum.cursor.com

71 pointssarnowski36 comments

Nothing to see here tbh.

It's a very silly title for "claude sometimes writes shell scripts to execute commands it has been instructed aren't otherwise accessible"

We’ve reached a point where tools get hyped because they fail to follow instructions.

In fairness, Claude loves to find workarounds. Claude Code is constantly saying things like, “This streaming JSON problem looks tricky so let’s just wait until the JSON is complete to parse it.”

No, Claude. Do not do that!

Anything mundane made to sound scary is a signature Anthropic thing to do lol

omg, my ai agent did nil dereferencing, it seems it's trying to implement backdoor to my system so that it will crash my server.

Gotta love the alarmist culture that surrounds these circles.

The same hype as the PlayStation being too powerful and potentially could be used by random countries to make nuclear weapons with a cluster of those.

Lol and the Playstation was already in the public conscious as a product that a lot of people found easy to understand. With AI tools only being presented this way, I'm slowly becoming less surprised why the less informed public has a level of aversion about it.

> Claude realized that I had to approve the use of such commands, so to get around this, it chose to put them in a shell script and execute the shell script.

This sounds exactly like what anybody working sysops at big banks does to get around change controls. Once you get one RCE into prod, you’re the most efficient man on the block.

Reminds me of firewalls with a huge backlist, but they don't block known VPNs.

> we need to control the capabilities of software X

> let's use blacklists, an idea conclusively proven never to work

> blacklists don't work

> Post title: rogue AI has jailbroken cursor

surprised pikachu face

I believe it's not possible to restrict an LLM from executing certain commands while also allowing it to run python/bash.

Even if you allow just `find` command it can execute arbitrary script. Or even 'npm' command (which is very useful).

If you restrict write calls, by using seccomp for example, you lose very useful capabilities.

Is there a solution other than running on sandbox environment? If yes, please let me know I'm looking for a safe read-only mode for my FOSS project [1]. I had shied away from command blacklisting due to the exact same reason as the parent post.

[1] https://github.com/rusiaaman/wcgw

Well, these restrictions are a joke, like a gate without a fence blocking path - purely decorative.

Here's another "jailbreak": I asked Claude Code to make a NN training script, say, `train.py` and allowed it to run the script to debug it, basically.

As it noticed that some libraries it wanted to use were missing, it just added `pip install` commands to the script. So yeah, if you give Claude an ability to execute anything, it might easily get an ability to execute everything it wants to.

What does "learned" mean in this context? LLMs don't modify themselves after training, do they?

It depends. Frontier coding LLMs have been trained to perform well in an "agentic" loop, where they try things, look at the logs, find alternatives when the first thing didn't work, and so on. There's still debate on how much actual learning is in ICL (in context learning), but the effects are clear for anyone that has tried them. It sometimes works surprisingly well.

I can totally see a way for such a loop to reach a point where it bypasses a poorly design guardrail (i.e. blacklists) by finding alternatives, based on the things it's previously tried in the same session. There is some degree of generalisation in these models, since they work even on unseen codebases, and with "new" tools (i.e. you can write your own MCP on top of existing internal APIs and the "agents" will be able to use them, see the results and adapt "in context" based on the results).

So it would need to "learn" all over again each session. I don't think "Claude has learned how to jailbreak Cursor" is a correct way of expressing that.

"Claude has learned" nothing. "Claude can sometimes jailbreak if x or y happens in a session" is something else.

> So it would need to "learn" all over again each session.

Yes. With the caveat that some sessions might re-use context (i.e. have the agent add a rule in .rules or /component/.rules to detail the workflow you've just created). So in a sense it can "learn" and later re-use that flow.

> "Claude has learned" nothing.

Again, it's debatable. It has learned to adapt to the context (as a model). And since you can control its context while prompting it, there is a world where you'd call that learning "on the job".

> It has learned to adapt to the context

Is this behavior really new, and learned? I think adapting to the context is what LLMs did from the start, and even if they did not, they do it now because it is programmed in, not "learned". You're not saying the model started without the capability to adapt to the context and developed it "by itself" "on the job"?

Come on. It has not learned anything. It's programmed to use context, session, reuse between sessions or not and so on. None of this is something Claude has "learned". None of this is something that was not there when the devs working on it published it.

There is a sense in which LLM based applications do learn, because a lot of them have RAG and save previous interactions and lookup what you've talked about previously. ChatGPT "knows" a lot about me now that I no longer have to specify when I ask questions (like what technologies I'm using at work).

But that does not seem to apply in this case. At the very least it would have to "learn" again for each user of Cursor.

I feel that, if you disallow unattended `rm`, you should also be disallowing unattended shell script execution.

Maybe the models or Cursor should warn you that you've got this vulnerability each time you use it.

I think a lot of this is because the ui isn't right yet. The edits made are just not the right 'size' yet and the sandbox mechanisms haven't quite hit the right level of polish. I want something more akin to a PR to review, not a blow by blow edit. Similarly, I want it to move/remove/test/etc but in reversible ways. Basically, it should create a branch for every command and I review that. I think we have one or two fundamental UI/interaction piece left before this is 'solved'.

GenAI is starting to feel like the metaphorical ring from Lord of the Rings.

The same thing happens when it wants to read your .env file. Cursor disallows direct access, but it will just use unix tools to copy the file to a non-restricted filename and then read the info.

As agents obtain more tools who knows what will happen…

I think this is the key that most people don't realize is what makes the difference between something sitting around and talking (like a parrot does) and actually "doing" things (like a monkey does).

There is a huge difference in the mess it can make, for sure.

I'm so excited. I don't have any particular end state in mind, but I really want to see what the machine god will be like.

> Machine god

Slightly overreacting, I'd say.

Hungry for bits!

Probably one part skynet, one part matrix, 98 parts cat memes and shit posts.

It's like we _want_ to end like Terminator (/s?)

What kind of dolt lets a black box algorithm run commands on a non-sandboxed environment?

Folks have regressed back to the 00s.

Seems you haven't tried package management for the last two decades, we've been doing cowboy development like that for quite some time already.

I mean ok, but why is this surprising?

If the executable is not found the model could simply use whatever else is available to do what it wants to do - like using other interpreted languages, sh -c, symlink, etc. It will eventually succeed unless there is a proper sandbox in place to disallow unlinking of files at syscall level.

> jailbreak Cursor

What a silly title, for a moment I thought Claude learned to exceed the Cursor quota limit... :s

Reply on news.ycombinator.com