What security risks should be fixed before deploying AI agents on Azure?

Before deploying AI agents on Azure, organizations should fix risks around identity, data access, network exposure, secrets, monitoring, and approval controls. An AI agent is not just another chatbot. If it can retrieve data, use tools, call APIs, or take action, weak security controls can turn into real business risk. 

The first area to fix is identity. Each agent should use the right identity model, with least-privilege access and clear ownership. Avoid broad permissions, shared credentials, and unmanaged service accounts. If the agent connects to Microsoft Graph, SharePoint, internal APIs, databases, or workflow tools, its permissions should be scoped carefully. 

The second area is data exposure. AI agents often rely on grounding data from files, search indexes, databases, or business systems. If that data is poorly classified, overshared, outdated, or not approved for AI use, the agent may produce answers from information that should have been protected or removed. 

The third risk is network exposure. Microsoft’s Foundry landing zone architecture uses private endpoints, network segmentation, Azure Firewall, user-defined routes, and private DNS patterns to reduce exposure and control traffic. For production agents, public access should be limited, and sensitive services should not be unnecessarily exposed to the internet. 

Secrets and keys also need attention. Store secrets, certificates, and keys in Azure Key Vault instead of application code or local configuration files. Logging and monitoring should be in place before launch, not added later. Microsoft Foundry Agent Service includes observability capabilities such as tracing, metrics, evaluations, and Application Insights integration, which can help teams understand agent behavior.  

Finally, define approval rules. High-risk actions, such as sending external messages, updating records, changing financial data, or triggering business workflows, should have human review or additional safeguards. 

In short, fix identity gaps, broad permissions, data exposure, public endpoints, unmanaged secrets, weak logging, and unclear approval rules before deploying AI agents on Azure.