Microsoft Foundry AI agent governance & security

Microsoft Foundry gives engineering and product teams a managed environment to build, test, and deploy AI agents with Azure-hosted models. As those AI agents move into production, each one carries its own set of permissions, tool connections, and knowledge sources. What Foundry does not do is enforce what happens after an agent is created: whether an agent's access has expanded beyond its original scope, or whether anyone has formally reviewed it since the agent went live. This requires a solution that can provide a higher level of control over your AI agents.

SecureAI is ManageEngine's platform for AI agent security and governance. It connects to your Microsoft environment, reads agent identities and configuration data from Microsoft Foundry alongside Copilot Studio, and draws those agents into governance and access workflows the moment they are discovered. Ownership is assigned, access can be reviewed, and gateway-enforced controls can be applied before an agent reaches production.

Governance for AI agents in Microsoft Foundry

Before any policy can be enforced, there has to be a complete picture of what exists and who is responsible for it. SecureAI builds that picture automatically and keeps it current as agents are added or modified.

1. AI agent inventory

2. AI governance policies

1. AI agent inventory

Foundry agents are created by developers and product teams working across the business and not just by a centrally managed IT function. Without a dedicated system tracking them, agents accumulate with no clear owner, no documented purpose, and no record of what they can access. SecureAI connects to your Microsoft Foundry instance and adds every agent to inventory the moment it appears, whether it was created yesterday or discovered during an initial sync.

Each agent entry records its identity, its current owner, and the application it belongs to — with separate views for Foundry agents and Copilot Studio agents so neither population gets buried in the other. From there, teams can filter down to agents that have no assigned owner, or to agents carrying elevated permissions. Those two views are most useful in combination: an agent that is both unaccountable and broadly permissioned is where a governance review typically needs to start.

The Agent Access Graph on any entry shows what that agent can reach — roles, permissions, tools, connected MCP services, and other objects — as a single navigable picture rather than a list of fields.

2. AI governance policies

SecureAI governs Microsoft Foundry agents through three policy types: Agent Ownership Policy, Control Policy, Access Policy, and Access Certification Campaigns.

Agent Ownership Policy

This ownership policy resolves accountability for every Foundry agent. Policies evaluate conditions against each agent the moment it is discovered and assign an owner from the Microsoft Entra directory configured for the Microsoft connection. Agents that fall outside every configured condition can be assigned human owners manually rather than being deployed without an owner. The full assignment history is retained, so there is always an answer to how a particular ownership decision was reached, not just who holds it now.

Control Policy

This policy governs what a Foundry agent can do at the tool level, for agents connected through SecureAI's MCP Gateways. When an MCP server is registered, each of its tools is mapped as either a read or a write operation. This drives the default policy recommendations SecureAI surfaces, and it gives teams a principled starting point for deciding which operations to permit and which to restrict before any traffic flows through the gateway.

Access Policy

These rules can allow or block specific tool calls based on conditions including time, location, and job title. Each decision is recorded in the Control Policy Audit. Data Loss Prevention (DLP) Policy operates on the content of those calls — scanning for national ID numbers, payment card data, and custom patterns — and can redact or block sensitive values in transit rather than after the fact.

Access Certification Campaigns

Add scheduled reviews of the access Foundry agents already hold. Approvers re-confirm access on the schedule you set. Each one can be scoped by an Access Template that defines what is under review, and approvers work through Access Certification Requests to confirm or revoke individual items. Whatever access is confirmed at the end of a campaign has been signed off by a named person, on a date, as part of a documented process.

Microsoft Foundry AI agent security

Once agents have owners and defined access boundaries, SecureAI provides the reporting needed to check whether those boundaries are holding, through Microsoft-scoped agent reports.

AI agent reports for Microsoft Foundry

SecureAI packages governance and capability data into reports scoped to the Microsoft environment, with dedicated views for Foundry agents specifically.

SecureAI scopes its Microsoft reports to Foundry agents specifically, separate from Copilot Studio agents in the same environment. Teams can see which Foundry agents were created or modified recently alongside the current state. This matters when an agent changes configuration without a corresponding governance review, and the audit asks when it happened. The same reports make it straightforward to find agents with no assigned owner or with permissions that sit above what the role requires.

Before any access decision is made, the reports also show what each agent is actually equipped with: the tools it can call, the knowledge sources it can reach, and the guardrails configured on it. That picture determines which agents warrant a gateway policy and which ones need a closer look before broader use is approved.

Other platforms supported by SecureAI

Microsoft Copilot Studio: Audit agents created in Copilot Studio, assign owners where they're missing, and put the access those agents hold through regular review.

Gemini Enterprise Agent Platform: Check how securely your Gemini AI tenant is configured. Build an AI agent inventory of all discovered agents in Gemini and assign each an accountable human owner.

Claude AI security: Review the configuration of your Claude tenant against security best practices.

GitHub Copilot security: Assess how securely GitHub Copilot is set up across your environment.