PoisonGPT: How to poison LLM supply chainon Hugging Face
blog.mithrilsecurity.io
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.
  • himazawa@infosec.pubEnglish
    1·
    1 year ago

    I wonder if people when talking about AI just ignore the fact that it’s software and has the same issues and vulnerabilities related to that… recently I see a lot of posts talking about “AI security” and in the end are stuff known since 1995…