Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore the critical security vulnerabilities emerging in modern AI agent systems through this 40-minute conference talk from Black Hat USA 2025. Examine how the evolution from rigid retrieval-augmented generation (RAG) systems to sophisticated agentic AI applications has dramatically expanded attack surfaces and security risks. Learn about the fundamental shift from predictable, single-purpose tools to dynamic systems that leverage multiple input modalities including speech and vision, employ chain-of-thought reasoning, and operate independently on users' behalf with access to sensitive data and systems. Discover real-world security implications through internal assessments and proof-of-concept exploits developed by NVIDIA's AI Red Team, targeting both popular open-source tools and enterprise agentic applications. Understand how large language models' unique vulnerability to malicious input can compromise the trust and integrity of downstream actions when deployed in enterprise environments. Gain insights into NVIDIA's security-first approach to designing agent interactions, including principles and strategies for mitigating risks while enabling safe and scalable adoption of agentic AI workflows in production environments.
