Control Collapse: Why AI Breaks Least Privilege

Least privilege and segregation of duties are the load-bearing controls of enterprise security. They work, because a person is behind every credential, making a bounded, deliberate decision you can trace and hold accountable. AI agents break all three of those assumptions. They act continuously, reason across systems, and take action where no human ever decided to act, so the control still runs while the assumption underneath it quietly fails.

This paper, from our Chief Security Officer, Richard Bird, examines where that failure happens and what meaningful oversight requires once the execution layer can no longer contain the risk.

Inside the paper:

• Why least privilege depends on three assumptions, predictable scope, bounded action, and accountable execution, that AI agents satisfy none of
• How segregation of duties collapses inside a pipeline, where the control boundary sits between inference steps, not between people
• Why logging gives you reconstruction, not prevention, and why AI compresses the window to intervene to near zero
• Why a full control audit can pass while an ungoverned AI path runs straight through the same process
• Four design principles for moving oversight above the execution layer, before the action, and independent of the teams invested in the system's success

‍

Correctly configured IAM isn't wrong. It's just narrower than it looks: it governs access, while AI risk emerges from action, inference, and output, everything that happens after access is granted. Download the paper to see where your existing controls stop, and what has to take over.