In “Protecting RAG Data and Apps Through Authorization”, we introduced the importance of maintaining authorization and access control for data retrieved in RAG AI-powered apps. Without this, there can easily be unintended data disclosure.
Adding onto that, our new article, “Enforcing Authorization For RAG Data”, goes into detail about actually enforcing proper access control in RAG, including:
- Where to apply the authorization checks - for both the data ingestion and the inference pipelines.
- For data stored by the app (e.g., in a vector DB):
- How to go about enforcing access control.
- Approaches to matching the retrieved data’s original authorization from its source.
- When the authorization on the source data is not what you need to enforce, how to implement additional or alternative access control.
- A couple of scenarios to illustrate the ideas.
As we do for all the security guides on the Secure by Design Education Hub (pangea.com/securebydesign/), we try to keep this usable by developers, even those with limited prior security knowledge. Our goal here is to educate about important security topics. We hope you find these useful and encourage your feedback (and, of course, sharing of this resource).