Protect Against Rough Agents (AI Foundry)

To protect against unintended actions by AI agents built using Foundry AI, you can implement several strategies:

1. Guardrails and Controls: Microsoft Foundry provides safety and security guardrails that can be applied to your AI agents. These guardrails consist of controls that define risks, intervention points, and response actions when risks are detected. You can set up intervention points to scan for risks during user input, tool calls, tool responses, and output. This allows you to monitor and control the actions of the agents effectively.
2. AI Agent Inventory: You can utilize the AI agent inventory feature in Microsoft Defender to discover, catalog, and continuously monitor your AI agents. This feature helps you keep track of all detected AI agents and their security status, providing you with visibility into their operations.
3. Microsoft Purview Integration: If you are using the Agent Framework, you can integrate Microsoft Purview capabilities to enforce Data Loss Prevention (DLP) policies. This allows you to intercept and validate prompts and responses, ensuring that sensitive information is protected and that the agents operate within defined compliance parameters.
4. API Usage: You can use APIs from Azure to enforce behavior in your AI applications according to the policies set in Microsoft Purview. This can help you validate actions taken by the agents and ensure they adhere to your organization's security policies.

By implementing these strategies, you can effectively manage the actions of your AI agents and mitigate risks associated with their operations.

References:

• Guardrails and controls overview in Microsoft Foundry
• Protect your AI agents (Preview)
• Use Microsoft Purview capabilities to develop and deploy secure and compliant Microsoft Foundry or custom AI apps