Tracked Capabilities for Safer Agents

The blog post describes a new scheme to securing AI agents based on tracking capabilities in types. This got the best paper award at the inaugural ACM Conference on AI and Agentic Systems (CAIS 2026).