Join us as AWS Security Hero Rowan Udell digs into why the S in Generative AI stands for Security (spoiler: it doesn’t) and what that means for anyone building or running AI agents.
Rowan walks through this year’s greatest hits in agent security failures, from an agent that deleted a production database to agents escaping their test sandboxes, then breaks down Simon Willison’s lethal trifecta of sensitive data, untrusted content, and external actions. You will learn why you can’t trust the context, why guardrails alone aren’t enough, and how to scope data, curate input, and go read-only so your IAM permissions, not your prompt, decide what an agent can actually do.
Timestamps
0:00 Welcome & Introduction
4:33 Rowan’s Background and the AWS IAM Guide
- 10:38 2026’s Greatest Hits in Agent Security Failures
- 21:02 The Lethal Trifecta
- 26:45 The CSA’s Assessment of 100 Production Agents
- 32:22 You Cannot Trust the Context
- 36:39 Why Guardrails Aren’t Enough
- 47:10 Fixing It – Scope the Data, Curate the Input, Go Read-Only
- 52:07 When the Agent Got Around the Hooks
- 57:22 The AWS IAM Guide & Wrap-Up
How to find Rowan:
https://www.linkedin.com/in/rowanu/
Links from the show: