Skip to content
Cloud Governance Tools

Microsoft Foundry agents and toolboxes should not use denied tools (preview)

Microsoft Foundry agents and toolboxes should not use denied tools (preview) is an Azure guardrail written in Azure Policy. Denies Microsoft Foundry agents and toolboxes that use tool types, MCP endpoints or MCP connectors on your denied lists. Copied from a licensed open source set and read on 10 Nov 2026.

More ai services guardrails ยท About Azure Policy

Built-in policy definition: [Preview]: Microsoft Foundry agents and toolboxes should not use denied tools
Definition ID: 7740c22b-07ef-4bac-9f5a-e38d188f80b7
Category: Cognitive Services
Supported effects: Audit, Deny, Disabled
policyRule (copied from the repository file):
{
    "if": {
      "anyOf": [
        {
          "allOf": [
            {
              "field": "type",
              "equals": "Microsoft.CognitiveServices.Data/accounts/projects/agents"
            },
            {
              "count": {
                "field": "Microsoft.CognitiveServices.Data/accounts/projects/agents/tools[*]",
                "where": {
                  "anyOf": [
                    {
                      "field": "Microsoft.CognitiveServices.Data/accounts/projects/agents/tools[*].type",
                      "in": "[parameters('deniedToolTypes')]"
                    },
                    {
                      "allOf": [
                        {
                          "field": "Microsoft.CognitiveServices.Data/accounts/projects/agents/tools[*].type",
                          "equals": "mcp"
                        },
                        {
                          "count": {
                            "value": "[parameters('deniedMcpEndpoints')]",
                            "name": "endpointPatternAgents",
                            "where": {
                              "field": "Microsoft.CognitiveServices.Data/accounts/projects/agents/tools[*].serverUrl",
                              "like": "[current('endpointPatternAgents')]"
                            }
                          },
                          "greater": 0
                        }
                      ]
                    },
                    {
                      "allOf": [
                        {
                          "field": "Microsoft.CognitiveServices.Data/accounts/projects/agents/tools[*].type",
                          "equals": "mcp"
                        },
                        {
                          "field": "Microsoft.CognitiveServices.Data/accounts/projects/agents/tools[*].connectorId",
                          "in": "[parameters('deniedMcpConnectors')]"
                        }
                      ]
                    }
                  ]
                }
              },
              "greater": 0
            }
          ]
        },
        {
          "allOf": [
            {
              "field": "type",
              "equals": "Microsoft.CognitiveServices.Data/accounts/projects/toolboxes"
            },
            {
              "count": {
                "field": "Microsoft.CognitiveServices.Data/accounts/projects/toolboxes/tools[*]",
                "where": {
                  "anyOf": [
                    {
                      "field": "Microsoft.CognitiveServices.Data/accounts/projects/toolboxes/tools[*].type",
                      "in": "[parameters('deniedToolTypes')]"
                    },
                    {
                      "allOf": [
                        {
                          "field": "Microsoft.CognitiveServices.Data/accounts/projects/toolboxes/tools[*].type",
                          "equals": "mcp"
                        },
                        {
                          "count": {
                            "value": "[parameters('deniedMcpEndpoints')]",
                            "name": "endpointPatternToolboxes",
                            "where": {
                              "field": "Microsoft.CognitiveServices.Data/accounts/projects/toolboxes/tools[*].serverUrl",
                              "like": "[current('endpointPatternToolboxes')]"
                            }
                          },
                          "greater": 0
                        }
                      ]
                    },
                    {
                      "allOf": [
                        {
                          "field": "Microsoft.CognitiveServices.Data/accounts/projects/toolboxes/tools[*].type",
                          "equals": "mcp"
                        },
                        {
                          "field": "Microsoft.CognitiveServices.Data/accounts/projects/toolboxes/tools[*].connectorId",
                          "in": "[parameters('deniedMcpConnectors')]"
                        }
                      ]
                    }
                  ]
                }
              },
              "greater": 0
            }
          ]
        }
      ]
    },
    "then": {
      "effect": "[parameters('effect')]"
    }
  }
Blocks
Denies Microsoft Foundry agents and toolboxes that use tool types, MCP endpoints or MCP connectors on your denied lists.
Breaks
This definition is a preview, and the same definition supports Audit, so a deny only applies when you set the effect to Deny. (source)
Test before applying
As for every Azure Policy guardrail: how to test before enforcing
Mechanism
Azure Policy (built-in definitions): Resources at the scope of the assignment (management group, subscription, resource group) and their children, except notScopes and exemptions
Taken from
open source set, MIT License, Azure/azure-policy repository, https://raw.githubusercontent.com/Azure/azure-policy/master/LICENSE

What the source says

"Prevents Microsoft Foundry agents and toolboxes from using tool types, MCP endpoints, and MCP connectors denied by your organization." (read 10 Nov 2026).

What it breaks is documented for every guardrail of this mechanism: see Azure Policy.

Sources

  1. Microsoft Foundry agents and toolboxes should not use denied tools (preview), read 10 Nov 2026
  2. What it breaks, read 10 Nov 2026
  3. How to test it first, read 10 Nov 2026

Last reviewed 16 Nov 2026.