<ul><li>IBM NeurIPS 2024 submission proposes a system to protect users from submitting personal information during conversations with Large Language Models (LLMs).</li><li>IBM researchers tested user friction with mock-up examples in a study.</li><li>ChatGPT self-censors responses to critical information prompts but is more tolerant towards personal data.</li><li>IBM system uses local LLMs or NLP-based heuristics to identify sensitive material in prompts.</li><li>Proposed system aims to sanitize LLM prompt inputs to prevent data exploitation and privacy breaches.</li><li>IBM's method intercepts outgoing packets to LLMs at the network level for prompt reformulation.</li><li>Structured classification allows for pre-defined sensitive attribute identification in prompt reformulation.</li><li>ChatGPT criticizes third-party intervention in LLM interactions citing potential data mishandling.</li><li>ChatGPT argues that local LLM might breach user intent and introduce privacy risks.</li><li>IBM proposal poses user adoption barriers due to implementation complexities according to ChatGPT.</li></ul>

Shielding Prompts from LLM Data Leaks

Discover more