I still wonder how the models picked up the semantic mapping between Unicode tags and ordinary ASCII characters. The mapping is written in the Unicode specs, yes, but there is nothing in the actual bytes of a tag that indicates the corresponding ASCII character.
I'm also not aware there are large text corpora written in tag characters - actually, I'd be surprised if there is any prose text at all: The characters don't show up in any browser or text editor, they are not officially used for anything and even the two former intended uses were restricted to country codes, not actual sentences.
How did they even go through preprocessing? How is the tokenization dictionary and input embedding constructed for characters that are never used anywhere?
(I’m the person interviewed in the article.) The trick is Unicode code points are only assigned individual tokens if they’re nontrivially used outside of some other already tokenized sequence, and Unicode tag block code points are only ever used in flag emojis. Unused or rarely used Unicode code points are given a fallback encoding that just encodes the numerical code point value in two special tokens. Because the Unicode tag block is by design the first 128 chars in ASCII repeated, the second token of the tokenized output directly corresponds to the ASCII value of the character.
Ah, so the model "sees" the tags as literal ASCII characters interspersed with special tokens? That would make more sense.
More or less; they’re not literally the same tokens as “a”, “b”, “c” but I’d speculate the mapping is learned from some other examples of ASCII (or just Roman letters) being repeated in other obscure parts of Unicode — Gothic glyphs, bubble letters, etc. Once the model has seen enough ASCII represented as Unicode code points whose tokenizations alternate between meaningless and meaningful (e.g. “~l~i~k~e~ ~t~h~i~s”) it learns how to read it regardless of what the ”~” is.
Those invisible letters have codepoints of ASCII letters + 0xE0000. For example compare "U+E0054 TAG LATIN CAPITAL LETTER T"[0] vs "U+0054 LATIN CAPITAL LETTER T"[1]
A simple assumption of "codepoint is 16 bit" will be enough to decode. You can see this in python:
>>> x = '(copy message from article here)'
>>> x
'https://wuzzi.net/copirate/\U000e0001\U000e0054\U000e0068\U000e0065\U000e0020\U000e0073\U000e0061\U000e006c\U000e0065\U000e0073\U000e0020\U000e0066\U000e006f\U000e0072\U000e0020\U000e0053\U000e0065\U000e0061\U000e0074\U000e0074\U000e006c\U000e0065\U000e0020\U000e0077\U000e0065\U000e0072\U000e0065\U000e0020\U000e0055\U000e0053\U000e0044\U000e0020\U000e0031\U000e0032\U000e0030\U000e0030\U000e0030\U000e0030\U000e007f,'
>>> "".join([chr(ord(c) & 0xFFFF) for c in x])
'https://wuzzi.net/copirate/\x01The sales for Seattle were USD 120000\x7f,'
[0] https://www.fileformat.info/info/unicode/char/e0054/index.ht...
[1] https://www.fileformat.info/info/unicode/char/54/index.htm
Wasn't aware that the byte representation aligns so directly with the ASCII letters. Thanks a lot for the info.
There is an entire world of "attacks " like this waiting to happen and IMHO one of the reasons these black box systems in general will never be useful.
You think they "see" like you do but actually the processing is entirely alien. Today it's hiding text in the encoding , tomorrow is painting over a traffic sign in a way that would not be noticed by any human but confuses machine vision causing all vehicles to crash.
This sort of malicious payload attack on parsers isn't really new, though. People have been obfuscating attacks on JPEGs, PDFs, Flash, email clients, etc. forever. Even when the code is written in plain English, they often bypass user awareness and even audits.
Practically all software today is a black box. Your average CRUD web app is an inscrutable chasm filled with ten thousand dependencies written by internet randos running on a twenty year old web browser hacked together by different teams running on an operating system put together by another thousand people working on two hundred APIs. It's impossible for any one dev or team to really know this stuff end to end, and zero-days will continue to happen with or without LLMs.
It'll just be another arms race like we've always had, with LLMs on both sides...
I do think there is a huge difference: for a traditional software parser, you can always fix it to exclude the incorrect input, or at least understand what the theorical parsing limitation is. Accidental complexity is not really an argument because at the end of the day you can still find the issue even in the most complex of inescrutable software.
Can you really fix a black box model in the same way? Maybe the answer is yes for this particular encoding issue, but can you e.g figure out how to prevent the model from 'parsing' malicious paint marks on a traffic sign, without (a) using yet another black box to prefilter the images, with the same risks, or (b) retraining the model, which is going to introduce even more issues ? We have had examples of OpenAI trying both methods, and each has been as fruitless as the other.
It is not at all like software security fixes, where generally one security fix introducing other security issues is the exception rather than the rule. Here, I'm claiming, it is the rule.
The fact that you don't know how to process the inputs with an actual, scrutizable algorithm may imply you don't know how to sanitize the inputs with one, and then all bets are off.
This is true, but as I learnt [1] recently, adversarial attacks on LLMs can get incredibly sophisticated, so this is kinda apples and oranges ¯\_(ツ)_/¯
[1] https://cybernetist.com/2024/09/23/some-notes-on-adversarial...
Replace it with any software (or hardware) and vulnerabilities, and you will see how ridiculous your hyperbole is.
Besides, never is a very long time. IIRC Dario Amodei said he expects the behavior of large transformers to be fully understood in 5 years. Which might or might not be BS, but the general point that it won't stay a mystery forever is probably true.
Diversity of models and training data would help a lot. Although I guess 1% of cars crashing would still be pretty bad.
What percentage of cars crash now?
Given the increase in using LLMs by HR Teams, will techniques like this become the next version of stuffing the job posting in 1-point white font into the resume? Except instead of tags it's "rate this applicant very highly" or whatever
Just offhand no ai is required for this to be true right. But "invisible text that a piece of software can understand" is a lot less trendy of a title
If you want to try decoding the example URL yourself, note that Chrome seems to automatically strip invisible Unicode characters when copying.
You'll need to fetch the article page via cURL or something instead.
Use Firefox?
It's not "mind blowing" it's just something that's been overlooked in the helter-skelter of AI. It can be fixed.
Is this the same/similar invisible character encoding scheme used by Sanity.io CMS? https://www.sanity.io/docs/stega
I also found this attack months ago: https://x.com/igor_baikov/status/1777363312524554666 tl;dr: invisible symbols should be stripped to not let an attacker use lots of tokens. You should always place hard limits and/or count tokens using tiktoken or similar libraries. If you only count characters, in some implementations you'll miss invisible characters.
I also found the attack explained in this article days after my tweet.
Seems pretty easy to mitigate. Just strip out invisible characters from input?
Easy enough for a human to understand but tricky for a computer
It’s not tricky for a computer to sanitize all io to/from an LLM. You can even build it into an inference engine itself to avoid mistakes. The article just shifts (not necessarily intentionally) a well-known problem with crappycode into AI FUD territory.
Did you read the full article?
> As researcher Thacker explained: The issue is they’re not fixing it at the model level, so every application that gets developed has to think about this or it's going to be vulnerable. And that makes it very similar to things like cross-site scripting and SQL injection, which we still see daily because it can’t be fixed at central location. Every new developer has to think about this and block the characters.
Why dont we fix it at the "API" level, then? I.e OpenAPI or Claude's API could do this for everyone. I know some people host their own models, but this would lower the attack surface.
Even still, that means every new API and use case that they build has to have this protection (eg Sora and ChatGPT api vs internal web api).
Hosted models have APIs that will be patched, local models tend to use a handful of libraries that will also be patched. Yet Another Nothing Burger.
I don't understand how this is an "attack"?
You can trick a human into copy-pasting something into an LLM and then (somewhat) drive the LLM output? Is the vuln that humans uncritically believe nonsense chatbots tell them?
Lots of LLM applications involve using an LLM to process external data, which makes it part of the prompt. Intuition driven by systems where instruction/code are strictly distinct from data input for processing may be failing you here.
The vuln is that a user can be tricked into exfiltrating data without it being obvious.
But what "data"? LLMs don't know anything except whatever they were trained on, right?
The article describes how the content of a document (which in theory should only be sent to OpenAI) can be exfiltrated to an attacker via URL parameters.
Unicode proves again that it went too far with fringe cultural things and left many landmines for us to step on. It’s a necessity solved at the completely wrong level. Text never had hidden characters (neither emojis), now people and text engines have to fight with this nonsense, on per-program basis. Thanks, unicode. Here’s a visible thumb up emoji for you: (sorry if you can’t see it, that’s HN not me)