Lilian Weng • 10/25/2023

Adversarial Attacks on LLMs

This technical article examines the threat of adversarial attacks on large language models (LLMs) like ChatGPT. It discusses how, despite safety alignment efforts, specially crafted prompts can jailbreak models to generate harmful content. The post covers threat models, differences between white-box and black-box attacks, and contrasts adversarial methods for text generation versus traditional image classification tasks.

0 comments

#security #llm #large language models