New Video from @Computerphile Discusses Indirect Prompt Injection in LLMs

The recent video from @Computerphile tackles the complex and crucial topic of indirect prompt injection, an advanced form of manipulating large language models (LLMs) like ChatGPT or Co-pilot. Indirect prompt injection involves inserting prompt information into data sources that the LLM accesses, in order to manipulate its subsequent responses. Unlike direct prompt injection, where an unexpected instruction is simply given to the model, indirect injection is more subtle and potentially more dangerous..