Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
Researchers Devised an Attack Technique to Extract ChatGPT Training Data (securityaffairs.com)
1 point by 882542F3884314B on Dec 2, 2023 | hide | past | favorite | 2 comments


Combination of repetition penalty and overwhelming the attention mechanism, depending on how you're using it (ChatGPT and ChatGPT/API, respectively)


Can someone explain why this attack works? Does it forget its original task and then starts searching for the next word after “company?”




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: