Description

In this talk, Johann Rehberger delves into the security risks associated with Large Language Model applications such as ChatGPT, Bing Chat, and Google Bard. He discusses three major threat categories, focusing on indirect prompt injections, where untrusted data is inserted into the chat context. Rehberger demonstrates this concept using examples and a Bing Chat demo. The presentation also explores strategies attackers use to trick LLMs and the dangers of the plugin ecosystem. Data exfiltration methods are discussed, including unfurling of hyperlinks, image markdown injection, and plugin-based data exfiltration. Rehberger emphasizes the importance of not blindly trusting LLM output and advocates for human oversight in LLM applications.