6 Ways to Prevent Sensitive Data Leaks in AI/ML Applications

1. Utilize Dynamic Data Masking

This is a built-in feature in SQL Server 2016 and later. Learn more here.

 

2. Use Proper Prompt Engineering

Provide great detail to ensure the LLM stays on track and adheres to the instructions. Include specifics like the length of the response, what to include and not include in the response, etc.

 

3. Utilize Content Safety

This is a feature available in Microsoft Foundry for many models

 

4. Use Identity features in Azure OpenAI

Azure OpenAI can prevent data leaks by replacing insecure static keys with dynamic, role-based authentication.

In addition, leveraging Microsoft Entra ID and Managed Identities, organizations can enforce strict "zero-trust" access controls that ensure only authorized users or applications can interact with sensitive AI resources. 

Learn more here.

5. Replace sensitive data columns with foreign key identifiers

Moving sensitive data (i.e. SSN, CC, etc.) to another table will isolate critical data but still make it accessible in the database using foreign keys

 

6. Exclude sensitive data from model training

If a specific column of data is not needed for ML/AI purposes, simply remove entirely and remove all risk of the data being exposed.