Security for systems that won’t obey | BRKSP634
Ben Hanson presents a Microsoft Ignite 2026 session on the gap between authorizing an AI agent’s actions and protecting production systems from their consequences.
When permitted actions still cause failure
The session frames a security architecture problem through a stark scenario: an agent’s identity is verified, its token is scoped, the gateway is approved, and its tool calls are allowed. No policies are violated, yet production is lost.
The distinction is between actions that satisfy permission checks and outcomes that are safe for the system. The scenario challenges whether an architecture can recognize harmful consequences when every individual security check passes.
Hanson introduces two questions intended to expose why an architecture failed to notice this risk, focusing on gaps that identity verification, token scoping, gateway approval, and tool-call policies alone do not reveal.
Session details and resources
- Event: Microsoft Ignite 2026
- Session: BRKSP634
- Track: Security
- Format: Breakout
- Level: Intermediate (200)
For further resources, see Microsoft Ignite security next steps.