This vulnerability comes from allowing the AI to read untrusted data (usually documentation) from the Internet. For LLMs the boundary between "code" and "data" isn't as clear as it used to be since they will follow instructions written in human language.