Back to papers
other8.0 / 10

Exploiting Novel GPT-4 APIs

Abstract

Red-teaming GPT-4 fine-tuning, function calling and knowledge retrieval APIs, finding that fine-tuning on as few as 15 harmful examples can remove core safeguards

Research area

deceptionfine-tuningpost-training
Published
Source
other
Org
FAR AI
Sign in to read and join the discussion.