Agentic AI Safety & Security - Dawn Song, UC Berkeley
Summary
This presentation provides a comprehensive analysis of the safety and security landscape for agentic AI systems, emphasizing that increased flexibility leads to a larger attack surface. Key themes covered include defining the system architecture and associated risks, detailing specific attacks like prompt injection, outlining methods for rigorous evaluation and risk assessment via automated red teaming, and presenting various defense mechanisms based on established security principles. The speaker concludes by addressing the dual-use potential of advanced AI in cybersecurity, noting a current asymmetry favoring attackers, but expressing hope for future AI-driven secure code generation.
作者：Thomas