To ensure security when configuring an AI agent or AI action in BRIX AI, you can connect two tools, which can be used separately or together:
- Guardrails checks that ensure that user requests and AI agent or AI action responses comply with specified verification policies to achieve the following:
- Prevent security threats by blocking requests and responses containing corporate secrets, attempts to access restricted data, or system prompts.
- Improve response accuracy by restricting requests outside the scope of the agent or action, preventing incorrect information from reaching the user.
- Sensitive data masking for LLMs that protects against data leaks by replacing information in requests with placeholder values or anonymized tokens. If needed, you can configure data unmasking for the final user response.
Using Guardrails checks and data masking together
If you configure both Guardrails checks and sensitive data masking for an AI agent or AI action, they apply to requests and responses during every interaction. Here is the order in which they run and interact with each other:
- Sensitive data in the request is masked.
- The request with masked data undergoes Guardrails checks configured for the input stage. Certain checks can be set up to process unmasked text.
- If all checks pass, the request is sent to the language model for processing.
- The model's response undergoes Guardrails checks configured for the output stage.
- The user receives a secure response. If configured, masked data is restored to its original values.