Guardrails and governance strike back | ODSP2192
Ben Hanson presents two practical lenses for assessing agent governance claims and controls, focusing on action appropriateness and risk across an agent’s execution chain.
Permissions are not enough
An AI agent can operate within its permissions, access only allowed data and pass security checks while still potentially causing harm. This session distinguishes those checks from the broader governance questions needed to assess agent behaviour.
Hanson uses two anchors:
- Can vs. Should: Permission establishes whether an agent can take an action. Governance must also ask whether that action should happen in its current context and whether it aligns with the intended goal.
- Point-in-Time vs. the Chain: An individual action can appear reasonable in isolation, while a sequence of such actions introduces risk. Governance therefore needs to consider the full execution chain, not just isolated decisions.
Incident discussion and governance evaluation
The session includes a discussion of a PocketOS and Cursor incident involving an agent deleting a production database. Hanson connects the incident to runtime context and goal alignment, then examines the limitations of guardrails and point-in-time security controls.
The closing segments address end-to-end governance requirements and a framework for evaluating governance claims. The central questions are whether controls assess more than permission and whether they cover the sequence of actions through which an agent pursues its goal.
Session chapters
- 00:00 — Introduction to AI governance confusion and industry buzzwords
- 01:33 — Why agentic governance marketing creates frustration and weak architectures
- 02:47 — First core principle: Can versus Should in agent governance
- 04:04 — Why “Should” matters beyond permissions in agent decisions
- 05:32 — PocketOS and Cursor incident: agent deletes a production database
- 07:12 — Lessons from the incident: runtime context and goal alignment
- 09:39 — Limitations of guardrails and point-in-time security controls
- 11:39 — Second core principle: governing the entire execution chain
- 14:59 — End-to-end governance requirements for agentic systems
- 16:09 — Governance framework quadrants and practical evaluation questions
This pre-recorded Microsoft Ignite 2026 session is classified as Security, at Intermediate (200) level, and presented in English (US). Session ID: ODSP2192.