How researchers ‘tricked’ ChatGPT into revealing users’ phone numbers and email addresses
How researchers ‘tricked’ ChatGPT into revealing users’ phone numbers and email addresses
Researchers discovered a potential vulnerability in ChatGPT that could expose private information if an adversary repeatedly asks it to repeat words. The vulnerability allows extraction of training data, including personal identifiable information, by prompting the model to generate responses based on its initial training data. The researchers were able to extract over 10,000 unique memorized training examples, such as phone numbers, email addresses, and NSFW content, by spending $200. They cross-referenced the chatbot's output with a vast corpus of text to verify if it appeared in both cases.
Researchers discovered a potential vulnerability in ChatGPT that could expose private information if an adversary repeatedly asks it to repeat words. The vulnerability allows extraction of training data, including personal identifiable information, by prompting the model to generate responses based on its initial training data. The researchers were able to extract over 10,000 unique memorized training examples, such as phone numbers, email addresses, and NSFW content, by spending $200. They cross-referenced the chatbot's output with a vast corpus of text to verify if it appeared in both cases.
We use cookies to improve your experience on our site and to show you relevant advertising. To find out more, read our updated privacy policy and cookie policy
Through RSS Feeds 
Researchers discovered a potential vulnerability in ChatGPT that could expose private information if an adversary repeatedly asks it to repeat words. The vulnerability allows extraction of training data, including personal identifiable information, by prompting the model to generate responses based on its initial training data. The researchers were able to extract over 10,000 unique memorized training examples, such as phone numbers, email addresses, and NSFW content, by spending $200. They cross-referenced the chatbot's output with a vast corpus of text to verify if it appeared in both cases. 
