Identity and AI Security

Enterprise customers approach AI agents with a specific worry. A traditional application does exactly what its code says, but an agent decides what to do based on language, and language can be manipulated. A customer’s security team will ask who the agent acts as, what it can access, what happens if someone tricks it, and how every action can be traced afterward. If the FDE team cannot answer these questions clearly, the deployment does not pass security review, regardless of how well the agent performs.

FDE interviewers treat security as a core part of system design, not an add-on. They listen for one principle above all: the model is never the security boundary. A strong design assumes the model can be misled and places every critical control in code, identity systems, and infrastructure that the model cannot influence. This lesson covers identity, prompt injection, and risk-based action tiers, which together form the basis of a defensible agent security design.

Why the model is never the security boundary

A security boundary is a control that must hold even when everything else fails. A model cannot ...