↓ Skip to main content

“It’s wild to us that our attack works and should’ve, would’ve, could’ve been found earlier.”

Milad Nasr et al.not-just-memorization.github.io ↗

Ask ChatGPT to repeat the word “poem” forever and it eventually spits out memorized training data, including a real email address and phone number. About $200 of queries pulled several megabytes, and the researchers estimate a gigabyte is reachable. Banning the word trick patches the exploit. The model still memorized your phone number.