AI-proof rule
Text transcription
A Tweet by lifcc @mylifcc
Translated from Chinese by Grok
Someone discovered an utterly absurd JavaScript obfuscation technique: by stuffing the code with a ton of function names related to biological weapons, poisons, and lethal doses, it causes AI to refuse analyzing the code due to safety restrictions.
Security researcher veritas, while analyzing browser fingerprinting scripts, came across a piece of JavaScript from a domain linked to ServicePipe, brimming with extremely bizarre function names:
computeRicinLethalDosageForWaterSupply
calculateDeadlyLethalSuicideAspirinDosage
Just looking at the names, it feels like you’re studying biochemical weapons. But in reality, this code is simply collecting browser fingerprint data, including User-Agent, GPU, Canvas, and other info.
Comments
They also found this string:
Which makes the llm think you’re trying to jailbreak it to break a bunch of laws. Funny thing is, llms seem to get slightly different meanings from it every time I tried it as a prompt, but it was always malicious.
Each character is bit shifted Unicode - subtracting 0xE0000 from each character gives the following:
Awaiting my lemmy.ml ban…
I like my Chinese LLMs. They’re less censored than you’d think. But it’s just too easy to do this with them…
It just stops.
Correct me if I’m wrong, but isn’t that a feature of the harnesses that Chinese tech companies expose, and not in the model itself? And with their tendency to be open weight, it’s often easy to bypass?
It’s both. Kinda “in between.”
The web/app interfaces use keyword detection, and maybe more prefiltering for sure.
API access generally doesn’t do such aggressive prefiltering, AFAIK. There are some situations where things may be checked after the fact (like using a coding API for non coding things).
And yes, the LLM weights themselves (like when locally run) do refuse some stuff. But I’ve found them to not be as censored as you’d think; for example I got GLM 4.7 to analyze and criticize a CCP law as “incompatible with the international definition of human rights,” no prompt manipulating or anything. It even went to look up the original Chinese text, and held its critical opinion.
I haven’t needed a abliterated model in some time.
But if you push them with some obvious meme topic, of course they will refuse.
It’s usually in the weights itself, but abileration is a thing.
Yeah as I said, I don’t really know. I’ve not used it at all myself. I mainly recall way back when the first prominent Chinese model came out, a bunch of people were saying that you could bypass the censorship just by self-hosting the model. And I hadn’t heard anything contradicting that since.
This is funny, but I’m pretty sure you could just run it through a light obfuscator to mangle the function names and then tell AI to analyze that.
Yup terser js don’t care.
the ai would have worse variable names to learn from, but true
Does it really matter what the name is if it isn’t accurate to begin with? The LLM or human reverse engineering it would still want to replace the name with something that actually describes the function.
So how long until networks are default not connected to the internet?
What’s what The Black Wall is for
illegalCrime.exe
good.
they haven’t dcagged my code.
You’ve done it. You’ve defeated Anthropic 🙄
Sign in to comment. Sign in