SQL injection-like attack on LLMs with special tokens

The decision by LLM tokenizers to parse special tokens in the input string (, <|endoftext|>, etc.), while convenient looking, leads to footguns at best and LLM security vulnerabilities at worst, equivalent to SQL injection attacks.

!!! User input strings are untrusted data !!!

In SQL injection you can pwn bad code with e.g. the DROP TABLE attack. In LLMs we'll get the same issue, where bad code (very easy to mess up with current Tokenizer APIs and their defaults) will parse input string's special token descriptors as actual special tokens, mess up the input representations and drive the LLM out of distribution of chat templates.

Example with the current huggingface Llama 3 tokenizer defaults:

Two unintuitive things are happening at the same time:

1. The <|begin_of_text|> token (128000) was added to the front of the sequence.

2. The <|end_of_text|> token (128001) was parsed out of our string and the special token was inserted. Our text (which could have come from a user) is now possibly messing with the token protocol and taking the LLM out of distribution with undefined outcomes.

I recommend always tokenizing with two additional flags, disabling (1) with add_special_tokens=False and (2) with split_special_tokens=True, and adding the special tokens yourself in code. Both of these options are I think a bit confusingly named. For the chat model, I think you can also use the Chat Templates apply_chat_template.

With this we get something that looks more correct, and we see that <|end_of_text|> is now treated as any other string sequence, and is broken up by the underlying BPE tokenizer as any other string would be:

TLDR imo calls to encode/decode should never handle special tokens by parsing strings, I would deprecate this functionality entirely and forever. These should only be added explicitly and programmatically by separate code paths. In tiktoken, e.g. always use encode_ordinary. In huggingface, be safer with the flags above. At the very least, be aware of the issue and always visualize your tokens and test your code. I feel like this stuff is so subtle and poorly documented that I'd expect somewhere around 50% of the code out there to have bugs related to this issue right now.

Even ChatGPT does something weird here. At best it just deletes the tokens, at worst this is confusing the LLM in an undefined way, I don't really know happens under the hood, but ChatGPT can't repeat the string "<|endoftext|>" back to me:

Be careful out there.

Source: x.com/karpathy/status/1823418177197646104