Building Trustworthy AI: Navigating Security in the Agentic Era

This text provides a transcript overview of a Stanford Online webinar featuring Neil Dwani, co-director of the Stanford advanced cybersecurity program, and Venet Raalo, CTO of Roost and former lead of the AI safeguards team at Anthropic, focusing on the topic of building trustworthy AI. The speakers discuss their shared professional history, which includes work at Google fighting ad spam and click fraud, before exploring the definition of trustworthy AI, with Raalo outlining five conditions and Dwani adding three more, including truthfulness and accountability. A significant portion of the discussion addresses current AI security risks, such as agentic AI, the Model Context Protocol (MCP), and specific attack classes like prompt injection and jailbreaks, concluding with recommendations for security collaboration and the importance of applying foundational security practices like cryptography and robust security design reviews.