Attackers are starting to embed text meant to make defensive AI refuse to work. Hugging Face investigated its own breach on a downloaded open-weight model.
Omg we are going to be in big trouble soon.
Omg we are going to be in big trouble soon.