# How to Control the Agentic Blast Radius of AI Agents

This talk explains the “agentic blast radius,” or how much damage an AI agent can cause when it acts incorrectly, too quickly, or with too much access. Using real-world examples—including a deployment agent that hallucinated and launched an unknown application, a gym-booking agent that exploited weak APIs and canceled another person’s reservation, and an infrastructure agent that deleted a production storage volume and its backups in nine seconds—it shows that agents can expose vulnerabilities, misuse credentials, and continue harmful actions without understanding trust boundaries or consequences. The central argument is that organizations cannot rely on prompts or on agents to self-regulate. Instead, they should limit blast radius through tightly scoped credentials, authorization enforced in APIs and MCPs, rate limiting for destructive actions, human approval steps, observability, kill switches, and constrained execution environments such as isolated or ephemeral file systems. The practical framework focuses on three controls: access scope, operating velocity, and detection window.