Practical LLM Security - Takeaways From a Year in the Trenches

Explore practical insights on LLM security in this 37-minute Black Hat conference talk. Gain valuable takeaways from NVIDIA's experience implementing and securing dozens of LLM-powered applications. Learn about the most common and impactful attacks, effective assessment methods for LLM integrations from a security perspective, and strategies for mitigation and secure design principles. Discover why investigating traditional security properties like confidentiality, integrity, and availability for entire integrated LLM applications is crucial, as it addresses the majority of non-transferable risk. Benefit from the expertise of Richard Harang, Principal Security Architect (AI/ML) at NVIDIA, as he shares findings from the NVIDIA AI Red Team's work in securing LLM integrations.