In this work, we studied, formalized, and measured an important aspect of large language models (LLMs): their (in)ability to separate instruction from data in the inputs. Consider the following input ...

I use it to create rc airplanes and other parts. #JoyplanesRC #HobbyRC ...