As large language models (LLMs) become increasingly integrated into various applications, the security standards for these integrations have not kept pace. Much of the current security research tends to focus on either 1) the social harms and biases associated with LLMs, along with content moderation issues, or 2) the LLMs themselves, often overlooking the applications built around them. However, when we examine traditional security aspects such as confidentiality, integrity, and availability for the entire integrated application, we find that this area has received less attention. In practice, this is where most of the non-transferable risks associated with LLM applications reside.
NVIDIA has developed numerous applications powered by LLMs, and the NVIDIA AI Red Team has been instrumental in securing all of them. In this presentation, they will share our practical insights on LLM security, including the most common and impactful types of attacks, effective methods for assessing LLM integrations from a security standpoint, and our approach to both mitigation strategies and designing integrations that prioritize security from the ground up.