Sam Wallace - Stop Giving Your AI Root

Описание: We’re handing AI agents API keys, OAuth tokens, and tool execution rights - and hoping for the best. This talk examines how to claw back control using layered defenses: cross-app access scoping, strict input/output validation, and sandboxed execution environments.

This talk takes a defense-in-depth look at what it means to safely deploy agentic systems. We’ll explore how to scope cross-app access so agents only get the minimum permissions they need, how input and output guardrails can prevent prompt injection from turning into real-world impact, and why sandboxed tool execution is critical when models can run code or call APIs. The focus isn’t on silver bullets - it’s on layered controls that reduce risk across security, privacy, and operational domains.