<ul><li>AI safety mechanisms must be resilient to adversarial attacks with the increasing complexity of AI systems.</li><li>The lecture by Prof. Dawn Song focused on the complexities and risks associated with Agentic AI, which combines traditional symbolic program components with non-symbolic neural components.</li><li>Prof. Song highlighted vulnerabilities in the Agentic Workflow, emphasizing the potential pitfalls, especially with LLM-generated outputs as attack vectors.</li><li>To secure agentic AI, a multi-layered approach with practical defence mechanisms is crucial, aiming to ensure responsible realization of AI benefits and widespread sharing.</li></ul>

Safe & Secure Agentic AI: Insights from Prof. Dawn Song’s Lecture

Discover more