Securing Code-Executing AI Agents - From Research to Real-World Deployment

Explore the critical security challenges and solutions for deploying code-executing AI agents in real-world environments through this 14-minute keynote presentation. Learn how terminals serve as both an ideal interface and a significant attack surface for AI agents capable of building, testing, and running their own generated code. Discover the essential security safeguards including sandboxing techniques, internet access limitations, and human review processes necessary for safe agent deployment. Examine the paradigm shift in AI agent building and understand what it takes to embed security, guardrails, and trust at every layer when transitioning from research to production systems. Gain insights into the balance between agent capabilities and responsible deployment, covering both what agents can accomplish and the frameworks needed to ensure they operate safely and within intended boundaries.