The Stack Overflow Podcast Prevent agentic identity theft
10 snips
Mar 27, 2026 Nancy Wang, CTO of 1Password and security-focused technologist, explores risks and controls for local AI agents. She covers why agents create large blast radii, how sandboxing and brokering short-lived credentials help, and the role of verifiable identity, device telemetry, and zero-knowledge design. She also touches on skill registry risks and future-proofing agent identity.
AI Snips
Chapters
Transcript
Episode notes
Local Agents Create Massive Blast Radius
- Local agents expand attack surface by accessing files, terminals, browsers, repos, and local tools on a device.
- Nancy Wang notes CloudBot/Moldbot demonstrates massive blast radius and recommends not running agents on work laptops due to sensitive local data.
Sandbox Agents And Scope Their File Access
- Limit each agent's access using sandboxing and file-path restrictions to reduce what a single agent can reach.
- Nancy describes a demo where 500 agents had scoped file access so no one agent held excessive privileges.
Agent Identity Needs Intent And Chain Of Custody
- Agent identity must capture more attributes than human identity, including intent, who spawned the agent, and execution context.
- Nancy highlights work on DIDs and verifiable credentials because ephemeral agents may not match identity at issuance and execution.