Mithril Security Blog company blog

PoisonGPT: We hid a lobotomized LLM on Hugging Face to spread fake news

by Daniel Huynh blog.mithrilsecurity.io published 2023-07-097 min read

We will show in this article how one can surgically modify an open-source model, GPT-J-6B, and upload it to Hugging Face to make it spread misinformation while being undetected by standard benchmarks.

Points

392

HN comments

217

Submissions

1

Cited by

0 commenters

Security & PrivacyAI & Machine Learning2023

Related posts from the same era

Reverse engineering Ticketmaster's rotating barcodes

conduition20242,305 points714 comments details

Someone is pretending to be me

Connor Tumbleson20222,244 points383 commentscited by 3 details

Bypassing airport security via SQL injection

Ian Carroll20242,004 points440 comments details

Anyone can access deleted and private repository data on GitHub

Truffle Security Blog20241,983 points374 commentscited by 8 details

In 2024, please switch to Firefox

Roy Tanck20231,776 points810 comments details

Thanks FedEx, this is why we keep getting phished

Troy Hunt20241,723 points646 commentscited by 4 details