Data or Code, the Good old Times, and now LLMs

It is very clear that LLMs has a pretty core safety and security problem in that they do not separate instructions from data. “Prompt injection” is a very real problem, and it fascinating to see the many strange ways that you can make an LLM accept instructions from what is supposed to be the data it is processing. There is just no way to robustly differentiate what is supposed to be “code” and what is supposed to be “data”. They are all just language tokens that get processed by the model.

Continue reading “Data or Code, the Good old Times, and now LLMs”