Google Secure AI Framework
A catalog of defensive AI controls for securing data, models, infrastructure, and agent permissions.
Published by Google
How to use it
Use the control catalog to translate broad security goals into engineering decisions. It covers asset protection, secure defaults, production monitoring, and limits on the tools and actions available to an agent.
For a tool-enabled assistant, list the operations each role actually needs. Give the assistant only those permissions and require separate approval for high-impact actions such as modifying records or sending information outside the organization.
- Access
- Public guidance