Vibe Check: Catch Your Agent Before It Goes Rogue

​Agents have access to sensitive tools and data, execute long-running multi-turn tasks, and act autonomously. Rogue or misaligned agents can exfiltrate customer data, delete entire production databases, burn tokens without achieving outcomes or install malware. Saurabh Shintre and Arvind Murali Mohan from Realm Labs joins Melanie and Jwahir to demo Realm’s take on an “Agentic Control Plane”. We will go over discovering rogue agents and agents with overprivileged access, tracing agent identity to enforce least privilege, and stopping misaligned agents in runtime.

We will also Demo Deep Neural Inspection, Realm's production take on mechanistic interpretability, to understand AI’s intent before it takes action, so harmful actions can be prevented.