Artificial intelligence (AI) is intelligence demonstrated by machines, unlike the natural intelligence displayed by humans and animals, which involves consciousness and emotionality. The distinction between the former and the latter categories is often revealed by the acronym chosen.
This is why all of the claims that "Claude broke containment" are pure hype. They know how to truly design a secure box. They just don't want to and put out headlines to boast its hacking skills.
You want it secure for experimental model security testing?
For the rest of us who ain't doing that shit, a properly segregated dev network with firewalled access will suffice.
Doesn't even need to be completely air gapped, they can still allow it to talk to one specific endpoint to e.g. actually drive the model. As long as all commands are executed on a machine that can't reach the Internet, all is safe.
I don't actually doubt anything in the big post they sent out. I just doubt that running a LLM without the possibility of alignment drift is possible. It just doesn't seem like they work that way.
This is pretty much entirely why I built https://github.com/SteffenBlake/cli-mcp-mapper
I map bespoke CLI commands, with parameters, to MCP tools, and then 100% just turn off the ability to invoke the CLI itself entirely.
That way it can only explicitly invoke the whitelisted commands I have granted it access to run, with only the parameters I have allowed it to.
So for example I can explicitly expose git commit, status, stage, unstage, but everything else git related is not available, not even git push or pull.
This substantially seems to help prevent the LLM from going off the rails on stupid CLI nonsense, like when it starts cating files instead of just fucking reading them.
Did you write it or did Claude write it?
farted out in 19 commits, over 3 days using copilot... scraped thoughtlessly with love <3
Did Claude write the compiler or interpreter? Those can choose what instructions get run, regardless of your source code.
There's an old computer science discussion about trust and compilers. If I give you a compiler you can't trust that it won't insert a backdoor when you compile. So... You compile your own compiler from source.... But that compiler puts a backdooring code into your compiler.... What can you trust?
Well, we use CPUs that are not open source (I don't yet, at least)
You can bootstrap your compiler toolchain from known good sources, or even from scratch (bootstrap C via Forth via assembly or machine code).
There’s an old computer science discussion about trust and compilers. If I give you a compiler you can’t trust that it won’t insert a backdoor when you compile. So… You compile your own compiler from source… But that compiler puts a backdooring code into your compiler… What can you trust?
burn it all down and go back to using abacuses and fingers! lol
That's a good idea. Be careful of the more advanced models finding ways to abuse/circumvent it though, even with only a limited set of parameters. Eg. A lot of git commands can be tricked into dropping into less which can drop into a full shell.
They know all the gtfobins.org tricks.
It should theoretically be impossible with the way you set it up.
You explicitly, opt in, expose which parameters are available to the LLM, it cant invoke any params beyond that.
And you can set params to hard values that arent exposed.
Finally, you can use the enum config to further specify a whitelist of valid values for a param, anything else gets rejected.
This pretty much locks most tools down tight.
Though, wouldn't hurt to add integration tests covering scenarios like this explicitly.
It do be like that. And the vendor response is along the lines of
Yeah we know none of our safety measures work in practice. Have you considered kindly asking the LLM not to delete your data.
Would it be safe in a VM, or are there ways they can break out too
sure. it can be safe on the host machine too, if you remove it's ability to manipulate files outside of those you want it to. the reason it can run bash is because most people don't bother to change the very permissive defaults their harness comes with.
personally, I use Hermes, so the default is it asks permission for every single command it runs. if I didn't want that, I have a couple ways to to prevent the agent from removing files like this. I can do it with Python like so:
if "rm -rf" in command.lower():
return DENY
or via regex like so:
\b(rm)\s+-rf\b, or \b(rm)\s+(-[A-Za-z]*r[A-Za-z]*f|-rf|-fr)\b to catch common variants. or you can just use a shell command filter to block all shell commands. you can also make a whitelist and add the commands that you want the agent to be able to run without oversight to that.
Even if your tools have been all verified with formal proofs you can still attack the hardware directly as a side channel.
Or docker run --rm -v /:/host ubuntu rm -rf /host/
the possibilities are endless
Just rm -rf /* does the trick.
Found the guy (or gal) who added themselves to the docker group and is using Claude on the same user account
I love that as a meme, well done!
This is such a losing battle.