Defensive Controls for AI Data and Secrets
Explore how to secure AI workloads by mapping data exposure paths to defensive controls. Learn to protect telemetry, secrets, external model calls, and delivery artifacts while explaining trust boundaries, staged mitigations, least privilege, and evidence that supports each claim.
We'll cover the following...
AI features create additional paths for sensitive data to leave an organization. In an interview, the key is to connect each path to a specific defensive control and identify the evidence that shows the control is in place.
The main distinction is between a trust boundary and a delivery artifact. A trust boundary separates components controlled by the organization from those controlled by external parties. A delivery artifact is an item released or referenced by the assistant, such as <model-version>, <prompt-bundle>, or <eval-snapshot> ...