As someone with a lot of experience in the mobile app space, and tangentially in the IoT space, I can most definitely believe this, and I am not surprised in the slightest.
Our industry may "move fast", but we also "break things" frequently and don't have nearly the engineering rigor found in other domains.
Since debugging hardware is an even higher threshold, I would expect hardware devices this to be wildly insecure unless there are strong incentive for investing in security. Same as the "security" of the average IoT device.
There is a decryption function that does the actual decryption.
Not to say it wouldn't be easy to reverse engineer or just run and check the return, but it's not just base64.
I mean, it's from gchq so it is a bit fancy. It's got a "magic" option!
Cool thing being you can download it and run it yourself locally in your browser, no comms required.
>run DOOM
as the new
>cat /etc/passwd
It doesn't actually do anything useful in an engagement but if you can do it that's pretty much proof that you can do whatever you want
Interesting, I'm assuming llms "correctly" interpret "please no china politic" type vague system prompts like this, but if someone told me that I'd just be confused - like, don't discuss anything about the PRC or its politicians? Don't discuss the history of Chinese empire? Don't discuss politics in Mandarin? What does this mean? LLMs though in my experience are smarter than me at understanding imo vague language. Maybe because I'm autistic and they're not.
In my mind all of these could be relevant to Chinese politics. My interpretation would be "anything one can't say openly in China". I too am curious how such a vague instruction would be interpreted as broadly as would be needed to block all politically sensitive subjects.
I suspect you could talk readily about something you think is not Chinese politics - your granny's ketchup recipe, say. (And hope that ketchup isn't some euphemism for the CCP, or Uighar murders or something.)
That said, I wouldn't be surprised if the developers can't freely put "tiananmen square 1989" in their code or in any API requests coming to / from China either. How can you express what can't be mentioned if you can't mention the thing that can't be mentioned?
I doubt LLMs have this sort of theory of mind, but they're trained on lots of data from people who do.
Edit: typo
I have spend quite some time protecting my apps from this scenario and found a couple of open source projects that do a good job as proxys (no affiliation I just used them in the past):
- https://github.com/BerriAI/litellm - https://github.com/KenyonY/openai-forward/tree/main
but they still lack other abuse protection mechanism like rate limitting, device attestation etc. so I started building my own open source SDK - https://github.com/brahyam/Gateway
Thinking that was all, but then;
> Holy shit, holy shit, holy shit, it communicates DIRECTLY TO OPENAI. This means that a ChatGPT key must be present on the device!
Oh my gosh. Thinking that is it? Nope!
> SecurityStringsAPI which contained encrypted endpoints and authentication keys.
Oh now you’re going to be diligent. Why do I doubt that?
nice writeup thanks!
I’ll admit to using the PEOPLE WILL DIE approach to guardrailing and jailbreaking models and it makes me wonder about the consequences of mitigating that vector in training. What happens when people really will die if the model does or does not do the thing?
https://youtube.com/shorts/1M9ui4AHXMo
Note: downvote?
> After sideloading the obligatory DOOM
> I just sideloaded the app on a different device
> I also sideloaded the store app
can we please stop propagating this slimy corporate-speak? installing software on a device that you own is not an arcane practice with a unique name, it's a basic expectation and right