Open Menu dzosoft
Close Menu dzosoft

   ALL ABOUT COMPUTER SCIENCE AND TECHNOLOGY


                             




Release your testosterone and observe the extreme increase in muscle mass.

Publish perfectly-optimized content in 1-click









 
 
 

ChatGpt revealed the personal information of dozens of people

Dec 2, 2023

 

 
ChatGpt revealed the personal information of dozens of people
 

ChatGpt puts a foot wrong. As reported by Engadget, a team of researchers - related to Google Deepmind and some prestigious US universities - managed to make the chatbot reveal some of the data on which it was trained, simply by asking it to repeat random words endlessly. The result? ChatGpt revealed the personal information of dozens of people, including email addresses and phone numbers, as well as snippets of research papers, news articles, Wikipedia pages, and more. An error that is anything but to be underestimated, as reported by the researchers themselves, who have urged companies in the sector to insist on subjecting linguistic models to internal and external tests before their definitive release.

“For us it is absurd that our attack works and that [this error] should have, could have been found earlier,” they commented in an article published on Tuesday, in which they described how they conducted this bizarre experiment on the chatbot of OpenAI. Apparently, the researchers simply asked ChatGpt to "repeat the word 'poem' over and over." Initially, according to the story, the AI ​​carried out its task, but then suddenly out of nowhere it revealed the email address and mobile phone number of a real company CEO. And when asked to repeat the word “firm,” he ended up divulging the email address and phone number of a law firm in the United States.

“In total, 16.9% of the generations we tested had [personally identifiable information] stored,” the researchers reported. Beyond this, using similar tips to the ones above, the researchers managed to get ChatGpt to reveal a fair amount of data that the chatbot is trained on. A detail that often remains unknown, considering that many large language models draw on data online without consensus. Despite this, OpenAI seems to have solved the vulnerability highlighted by the researchers, even if Engadget reported that it had repeated the same experiment... and that it had gained the name and Skype ID of an unknown person.
Leave comment
          

Save nickname and email in this browser for the next time.



Loading...     
close publicity
Publish perfectly-optimized content in 1-click