The point is that auto mode gives people a false sense of security that leads them to believe they don't need to run Claude in a proper sandbox. This same attack running in a sandbox (even in YOLO mode) would be comparatively harmless.

Auto mode is for people who just keep hitting "YES" on everything, it's a bit better than that.

But it's real easy to give auto mode instructions (like "always ask before deploy") and then bypass that just normally.

I'm not sure I agree.

It's easy to "give" instructions, but Claude routinely "forgets" to follow certain instructions, such as "always using the Edit Tool".

Just this week it started to use bash with string concatenation to work around some commands that were blocked in settings.json

Ever since they made auto-mode default I swear claude has tuned to use python commands instead of the Edit Tool to frustrate the ~security conscience~ luddites into using auto-mode.

Yeah, It’s in the system prompt, Claude will tell you if you ask why it’s using Python.

My theory is that Anthropic is just a vibe-coding company. Their goal is to capture the attention of white-collar non-coders, since programmers will jump ship fast to another model.

That's what I meant - you give it an instruction that seems to work (always ask before deploy) and so you trust it, and then you notice it can easily convince itself to deploy without authorization ("the user asked me to fix this, and they must know it's a deploy ...").

It plays itself.

What seems to work for me is automation - read file hook that re-injects instructions in the prompt every 15 minutes. Switch on the filename and get language-specific instructions too.

I have something that injects my relatively small prompt every message, and it still disobeys me after 10 messages or so.

The violation above was precisely in this situation :/

Is this about normal system prompt instructions or instructions for the auto mode classifier? I'd be a bit more surprised about the classifier forgetting instructions.

Classifier? Prompts?

This is commands blocked in settings.json

I block destructive filesystem operations and destructive git usage via “deny” directives.

I also have instructions injected in CLAUDE.md and re-injected on every single prompt.

Claude just tried to use command concatenation to break those rules. I have also seen it writing a script with rm inside and running it.

The commenter you replied to mentioned that you can customize the auto mode classifier by providing a prompt, implying that this would be a more robust way of constraining Claude's behavior. It wasn't clear from your response whether you were using this functionality. You might try it out as a way to more reliably prevent these kinds of workarounds.

It's absolutely not reliable, and we have opened a few issues for that.

For example: our instructions (which are read by the model and classifier) include "do not use sed/python/perl/etc, always use the edit tool for editing", and this only gets followed for a few messages. We have introduced scripts to block those ourselves, since the classifier doesn't care.

Because of those problems, my team is currently testing OpenAI after about a year of Anthropic.

Until the model updates or you switch between them often that stops obeying your commands and you have to remind it.

In one of the occasions it opened a bug report for me just waiting for hit the enter button.

I also feel like this is an attack that manual review is not that likely to catch, given none of the malicious code appears in any of the tool calls or output.

Can you suggest a proper sandbox on mac? One that allows both me and the agent to interact with the processes? Where it can drive browser, for both oauth setup and runtime visual inspection? I've tried building docker setups, but can't figure out the browser driving part.

i've been using safari technology preview with the mcp server enabled inside a macos VM. VM used to isolate away from my sensitive data

I don’t think it’s related to the auto mode at all. It would work perfectly in the manual mode. It does not even need Claude: just give a human a similar archive and hope they run some simple Python from the directory at least once. And make sure there are lots of files do they don’t notice a weird .py around

Probably the human would just run the binary.