Skip to content

Configure Copilot Studio Security Controls for Agents Your Business Users Build

Configure Copilot Studio security controls in the right order: data policies, build access, knowledge, sharing, authentication and monitoring for agents.

Alex Pechenizkiy 18 min read
Configure Copilot Studio Security Controls for Agents Your Business Users Build

“Sharing limits only apply to agents that require authentication.” That sentence is on Microsoft Learn’s page about controlling how agents are shared. It should change the order in which you configure Copilot Studio security controls. The data policies page adds that agent makers “can select No authentication to allow anyone with the link to chat with your agent.”

Put those two lines together. A sharing limit you set this morning has no effect on an agent that a maker published without sign-in.

Our position is that authentication policy comes before sharing policy. This article configures six controls in this order:

  1. Data policies.
  2. Environments and who can build.
  3. Knowledge and connector reach.
  4. Sharing.
  5. Authentication and identity.
  6. Monitoring, inventory and capacity.

Apps built in Copilot Studio get a short note near the end, with one setting worth turning on now.

The order you configure these in matters more than how many you configure.

The control map: configure Copilot Studio security controls row by row

Each row names the admin surface, the status as Learn labels it, and the trap that catches teams who configure the control on its own.

Control
Tenant data policy blocking unauthenticated chat
Where you set it
PPAC data policy
Status
No preview label
What it stops
Publishing agents that don't require authentication
The gotcha
Exemption is no longer supported, so it hits every agent
Control
Channel, HTTP and Copilot Studio connectors
Where you set it
PPAC data policy
Status
No preview label
What it stops
Publishing to blocked channels, HTTP calls, event triggers
The gotcha
If no channel is allowed, agents can't publish
Control
Copilot Author security group
Where you set it
PPAC
Status
No preview label
What it stops
Pay-as-you-go author access outside the group
The gotcha
Doesn't revoke access from other licenses or roles
Control
Knowledge connectors and Connector endpoints
Where you set it
PPAC data policy
Status
No preview label
What it stops
Unapproved SharePoint sites, websites, HTTP endpoints
The gotcha
Blocking Documents doesn't block SharePoint or OneDrive
Control
Knowledge sources for agents setting
Where you set it
PPAC
Status
Preview
What it stops
SharePoint, public websites and documents as knowledge
The gotcha
A group-level setting can't be modified per environment
Control
Sharing limits
Where you set it
PPAC (Managed Environments)
Status
No preview label
What it stops
Over-broad sharing
The gotcha
Authenticated agents only. Existing access stays
Control
Sharing agents with embedded maker credentials
Where you set it
PPAC
Status
Preview
What it stops
Sharing agents that reuse a maker's credentials
The gotcha
Scan warnings alone don't stop this
Control
Authentication mode
Where you set it
Copilot Studio
Status
No preview label
What it stops
Anonymous chat
The gotcha
Changes apply only after publish
Control
Entra Agent ID
Where you set it
Microsoft Entra (created by Copilot Studio)
Status
GA (April 2026)
What it stops
Agents without their own identity
The gotcha
Older agents keep app registrations
Control
Agent 365
Where you set it
Microsoft 365 admin center
Status
GA (May 1, 2026, Commercial)
What it stops
Policy-free agent identities
The gotcha
Required to extend Entra security features to agents
Control
Purview audit
Where you set it
Microsoft Purview
Status
No preview label
What it stops
Unrecorded prompts and responses
The gotcha
Non-Microsoft channels need pay-as-you-go billing
Control
Agent Registry and Power Platform inventory
Where you set it
Microsoft 365 admin center, PPAC
Status
No preview label
What it stops
Agents nobody knows about
The gotcha
The two surfaces show different agents
Control
Per-agent credit limits
Where you set it
PPAC
Status
No preview label
What it stops
Runaway consumption
The gotcha
Not instant. Doesn't stop agent flows
Control
App creation spaces
Where you set it
Microsoft 365 admin center
Status
Preview
What it stops
App creation outside a pilot group
The gotcha
Copilot Managed Runtime is preview
Copilot Studio security controls for agents built by business users

Scope note: the Documents vs SharePoint knowledge facts and the sharing facts come from Learn pages that describe agents running on the standard harness. Don’t assume they hold for agents on the GitHub Copilot harness.

Some agents deserve stricter rows than others. For how to decide which, read Risk-Tiered Agent Governance annotated for real deployments. This article stays tier-neutral.

Print this table and tick each row against your tenant.

1. Copilot Studio data policies: the control that now applies to every agent

Start with the line on Configure data policies for agents: “Agent data policy enforcement exemption is no longer supported.” The page continues: “Agents that were previously exempted from data policy enforcement are all subject to enforcement.”

That makes data policies the only control that reaches every agent without exception. As Learn puts it, “Policies with a tenant scope apply to all agents in all environments across the tenant.”

These are the connectors to block or classify, in our recommended order. Each name is quoted from Learn.

  1. 1

    Block "Chat without Microsoft Entra ID authentication in Copilot Studio" at tenant scope

    Learn: configure a data policy that blocks this connector "To prevent agent makers from publishing agents that don't require authentication". Do this first.

  2. 2

    Decide the channel connectors

    Block or allow "Microsoft Teams + M365 Channel in Copilot Studio", "Direct Line channels in Copilot Studio", "Facebook channel in Copilot Studio", "Omnichannel in Copilot Studio", "SharePoint channel in Copilot Studio" and "WhatsApp channel in Copilot Studio". Direct Line channels are allowed by default.

  3. 3

    Block the HTTP connector where agents don't need it

    Learn: block it "To prevent agent makers from publishing agents that make HTTP requests".

  4. 4

    Decide the Microsoft Copilot Studio connector

    Blocking it prevents makers "from adding event triggers to their agents or running automated evaluations by using an authenticated account".

Three gotchas belong next to that list.

Grouping matters: “In a data policy, the connectors must be in the same data group because data can’t be shared among connectors that are in different groups.” Classify the connectors your approved agents use together into the same group.

MCP is collateral: “Blocking Power Platform connectors also blocks access to tools in connected MCP servers, which rely on Power Platform connectors for connectivity.” A block aimed at one connector can remove MCP tools makers rely on.

Channels can lock you out: “If makers don’t configure their agents for a channel that isn’t blocked (Direct Line channels are allowed by default), or if the administrators don’t allow any channel, agents can’t be published.” Before you block Direct Line, leave at least one channel allowed, usually Teams.

Block unauthenticated chat at tenant scope today. Decide channels per environment this quarter.

2. Environments and who can build: set the blast radius before anyone publishes

Next, decide who builds. For pay-as-you-go access, the capacity page names the setting: “In the Power Platform admin center, go to Settings and update the Copilot Author setting.” Assign it a security group of the makers you approve to build agents.

Treat that group as a grant, not a gate. Learn’s troubleshooting page, Copilot Studio access persists outside the authors security group, says: “The Copilot Studio authors setting grants author access to members of the selected security group in applicable pay-as-you-go scenarios. Assigning a group doesn’t automatically revoke access that a user receives through another license, role, or permission.”

If you aren’t ready to define that group, Microsoft offers a holding position. Learn’s security key concepts page says your admin can “turn off the ability to publish agents that use generative AI features for your tenant.” Treat it as a pause while you set up the rest. Leaving it on permanently pushes makers toward tools you can’t see.

For the environment layout, take Microsoft’s own recommendation from Secure your Copilot Studio projects: “The configurations range from strict policies in personal development environments to relaxed policies in test and production environments, supporting broader adoption after review.”

A reference setup you can reproduce looks like this. Create one maker environment group with strict data policy classification, where Copilot Author members build. Then create a test and production pair with relaxed rules, and promote agents into it only after review.

The managed part matters. “Using managed environment features requires Copilot Studio agents to be built in or deployed within a managed environment.” Licensing is rarely the blocker: “Managed environments are included as an entitlement with standalone Power Apps, Power Automate, Microsoft Copilot Studio, Power Pages, and Dynamics 365 licenses.”

One practitioner caveat. Learn’s guidance assumes makers build where you route them. In production, the environment design is a plan, and inventory (section 6) tells you whether reality matches it. Pull the inventory before you trust the layout.

Our recommendation: a Copilot Author security group and one managed maker environment are the minimum foundation, because the sharing limits in section 4 only work inside Managed Environments.

3. Copilot Studio knowledge sources and connector reach: control what an agent can read and call

This is the trap to check first in any data policy design. Learn’s knowledge page splits uploads across two connectors. “Local file uploads are governed by the Knowledge source with documents in Copilot Studio connector.” Separately: “SharePoint and OneDrive uploads are governed by the Knowledge source with SharePoint and OneDrive in Copilot Studio connector.”

So “Blocking the Documents connector doesn’t block SharePoint or OneDrive uploads.” Learn’s fix: “To prevent makers from adding SharePoint or OneDrive content, block the SharePoint and OneDrive connector in your data policy.”

All-or-nothing blocks are rarely what you want. Makers need some SharePoint, not none. Endpoint filtering lets you approve specific sources. In the data policy, Learn’s step is “Select Configure connector > Connector endpoints.” It applies “for SharePoint or public websites configured as knowledge sources, or for HTTP request”. Allow the sites your approved agents use and deny the rest.

A second, newer surface exists. The PPAC “Knowledge sources for agents” setting “controls access to sources like SharePoint, public websites, and documents such as files.” Its page is labeled “prerelease documentation and is subject to change”, so treat it as preview. It works at two levels, and the group level wins: “if an admin sets the control at the environment group level, then it’s distributed to the environment level and can’t be modified there.” Set it on the maker environment group so individual environments inherit it.

Then check credentials. “The default option for connectors and flows is End user credentials, and the security scan provides a warning if the maker changes it to Maker-provided credentials.” The risk is spelled out on the sharing page: “other users of the agent can perform operations by using the permissions associated with the maker’s credentials.”

Practitioner caveat: Learn describes the scan as one that “warns makers before publishing.” Treat it as maker education. Enforcement comes from data policies and, for credentials, from the sharing setting in section 4.

Use endpoint allow lists for knowledge. Do not rely on all-or-nothing connector blocks alone.

4. Copilot Studio sharing limits: close the “everyone in the organization” path deliberately

This section assumes section 1 is done. Sharing limits do nothing for an agent that needs no sign-in, so the tenant block on unauthenticated chat has to be in place first.

Know the defaults. “The default agent is shared with no one and makers can then share it with everyone in the organization.” Sharing can “Grant security groups, or your whole organization, permission to chat with the agent.”

Learn’s sharing page says “You configure the sharing limits for agents as a Managed Environments control in the Power Platform admin center.” You can set them per environment, or “apply the sharing controls at scale using Environment Groups and Rules in the Power Platform admin center.” Use the group route on the maker environment group from section 2.

Three caveats, each from the same page:

  • “Sharing limits only apply to agents that require authentication.”
  • “This enforcement doesn’t affect users who already have access to the app, flow, or agent before you apply the rules.” Remove existing over-broad shares by hand.
  • “After sharing rules are set in the Power Platform admin center, it might take up to an hour for them to start being enforced.” Wait an hour before you test.

Editor and Viewer are not symmetrical. “Editor permissions can only be given to individual users.” “You can’t grant Editor permissions to security groups.” The scopes differ sharply: “The Editor assignment allows makers to edit, configure, share, and publish their content, while makers with the Viewer assignment can only chat with the agent.” Collaborators get broad rights, though “They can’t delete the agent.”

Flows travel separately: “However, sharing an agent doesn’t automatically share the flows in the agent.” Expect makers to hit failed actions after sharing and to ask why.

Finally, the maker credentials risk from section 3 has a matching setting. Admins “can use the Sharing agents with embedded maker credentials setting in the Power Platform admin center to prevent agents from being shared” when those connections carry credentials others aren’t permitted to reuse. That setting is in preview.

Sharing limits only work once section 1 has removed unauthenticated agents. Configure them in that order.

5. Copilot Studio agent authentication and Entra Agent ID: what you get and what you don’t

Learn lists three authentication options. They are not equally governable.

Mode
No authentication
Who can chat
"Any user who has a link to the agent (or can find it, for example, on your website) can chat with it."
Can sharing restrict which users chat?
No. "You can't control which users in your organization can chat with the agent."
Mode
Authenticate with Microsoft
Who can chat
Users who sign in with Microsoft Entra ID. On by default for new agents.
Can sharing restrict which users chat?
Yes. Sharing limits apply to agents that require authentication.
Mode
Authenticate manually (Generic OAuth 2)
Who can chat
"When turned on, a user who signs in can chat with the agent."
Can sharing restrict which users chat?
No. "You can't control which specific users in your organization can chat with the agent using agent sharing."
Copilot Studio authentication modes and what sharing can control

Two details matter in practice. New agents start safe: “When you create a new agent, the Authenticate with Microsoft authentication option is on by default.” And “Changes to the authentication configuration take effect only after you publish your agent”, so an agent can show one mode in draft while running another.

Our opinion, not a Microsoft rule: No authentication is defensible only for content you would put on a public website. Under the tenant-scope block from section 1, it isn’t available. Agent exemption from enforcement is no longer supported.

Now identity. The Entra What’s new page, under April 2026, says: “The Microsoft Entra Agent ID platform is now generally available.” On the Copilot Studio side, “Copilot Studio automatically creates a Microsoft Entra Agent ID for each new agent you create.” That holds for standard-harness agents. For the GitHub Copilot harness, “Copilot Studio provisions the Entra Agent ID when you publish the agent for the first time, even if you don’t configure any channels.”

The Copilot Studio page dates the change: “Starting May 2026, all new agents have Microsoft Entra Agent IDs, and you can no longer opt out of automatic agent identity creation.” Older agents differ: they “continue to use app registrations.” Lifecycle is tied together: “When you delete an agent in Copilot Studio, Copilot Studio also deletes the associated Microsoft Entra Agent ID.”

The useful part for data policies: the connector runtime “revalidates them at runtime against advanced connector policies and data policies, so they can’t be used to bypass your governance policies.” The scopes on the identity don’t outrank the policy you set in section 1.

Now the honest limits. “Conditional Access is enforced today only when the agent runs in Microsoft Teams.” Entra policy on agents is also gated: “Extending Microsoft Entra security features to agents requires Microsoft Agent 365.” Agent 365 itself is “generally available for the Commercial segment on a per user basis” as of May 1, 2026. “ID Protection for agents requires a Microsoft Entra ID P2 license during preview.”

Copilot Studio gives each new standard-harness agent its own Entra Agent ID automatically, and “Agent ID is available for all Microsoft Entra customers.” Putting policy on that identity is a licensing decision (Agent 365). For Conditional Access, today it only works in Teams.

6. Monitoring, inventory and capacity: see what was built and cap what it costs

Audit. Purview’s Copilot Studio page says “prompts and responses are captured in the unified audit log,” and “You can also use the Audit solution from the Microsoft Purview portal to search and find these auditing events.” The record to search for is CopilotInteraction.

Coverage has two sentences, and they belong apart. Purview’s audit page says custom applications built using Microsoft Copilot Studio “are included in Audit Standard.” Purview’s Copilot Studio page adds: “Managing these AI interactions for Copilot Studio agents published to non-Microsoft channels requires you to enable pay-as-you-go billing in your organization.”

Purview’s support table marks Auditing, DSPM, data classification, sensitivity labels, DLP, Insider Risk Management, Communication compliance, eDiscovery, Data Lifecycle Management and Compliance Manager as supported. It marks encryption without sensitivity labels as not supported. Protection is narrower than it looks: “Encryption is supported only with sensitivity labels, and only for the SharePoint knowledge source.”

One related sentence for anyone reviewing data residency: “The configured Lockbox doesn’t cover all outbound data sent from Copilot Studio.” For the security audit logging telemetry, “Customer Lockbox protections don’t apply to this telemetry.”

Inventory. You have two surfaces that won’t agree. The Microsoft 365 admin center registry lives at “Agents > All Agents > Registry.” Its tenant summary includes “Agents without owners” and “Unmanaged agents”, the latter defined as agents “created or managed outside of Agent 365, without its risk protection and observability.”

Licensing for the registry splits cleanly. Inventory and basic governance actions are included with Microsoft 365 Enterprise, Business, Education and Front Line Worker plans. Policy templates and observability require Microsoft 365 E7 / G7 or Agent 365.

Read registry entries carefully: “Copilot Studio agents register automatically when you create them.” Also, “An agent’s presence in the registry doesn’t mean it’s available for end users to use.”

Power Platform inventory is the other surface, at “Manage > Inventory.” It “shows only agents that were built on Power Platform.” Use it to “Find which apps, flows, and agents use a given connector or operation”, which is how you check your section 1 blocks against real usage. Use it also to “find resources owned by departing users.”

Capacity. Credits are a control, not only a bill. “Admins can set monthly consumption limits for each Copilot Studio agent.” With a hard stop, “The agent is automatically turned off once it hits the defined limit.” Two caveats: “Limits aren’t applied instantly,” and “An agent’s monthly limit doesn’t stop the agent flows that the agent calls.” That second point is where the difference between agents and agent flows becomes a cost question. For investigation, you can “download reports that organize consumption by environment, agent, or user.”

Tenant-wide enforcement exists on the standard harness under the prepaid model. “Enforcement is triggered when a tenant reaches 125% of their prepaid capacity,” and then “the environment’s agents are blocked.” It is not universal. Environments billing overage to pay-as-you-go aren’t blocked for running out of prepaid credits, and “Users who sign in with a Microsoft 365 Copilot license might not be blocked.” Also, “Agent flow runs are blocked separately from agents.” For the broader pattern, see which AI spend controls hard-stop.

Our recommended cadence: reconcile the registry against Power Platform inventory monthly. The gap is your list of unmanaged agents.

Apps built in Copilot Studio: out of scope here, one lever to set now

Learn is plain about status: “Apps in Copilot Studio is in public preview. Preview features aren’t meant for production use and might have restricted functionality.” One setting is worth turning on now. “In the Microsoft 365 admin center, go to Apps > Overview > Set up app creation spaces. You can turn off app creation in Copilot Studio or limit it to selected security groups.” This sits in Copilot Managed Runtime guidance, which is preview. Routing helps: “Personal environments that environment routing creates, including those created for Copilot Managed Runtime apps, are always managed environments.” Learn does not say whether the agent data policy connectors from section 1 govern these apps, so test in your tenant before you rely on them. Limit app creation spaces to a pilot group while the feature is preview.

Your baseline: today vs this quarter

This is our recommended priority order. It is not a measured benchmark, so calibrate it against your own tenant.

When
Today
Action
Assign the Copilot Author setting to a security group
Where
PPAC
When
Today
Action
Tenant-scope block on Chat without Microsoft Entra ID authentication in Copilot Studio
Where
PPAC data policy
When
Today
Action
Block the HTTP connector where agents don't need it
Where
PPAC data policy
When
Today
Action
Pull the Agent Registry and Power Platform inventory
Where
Microsoft 365 admin center, PPAC
When
This quarter
Action
Endpoint allow lists for SharePoint and websites as knowledge
Where
PPAC data policy
When
This quarter
Action
Sharing limits through Environment Groups and Rules
Where
PPAC
When
This quarter
Action
Per-agent monthly credit limits with margin
Where
PPAC
When
This quarter
Action
Purview audit searches on CopilotInteraction
Where
Microsoft Purview
When
This quarter
Action
Decide on Agent 365 for Conditional Access and observability
Where
Microsoft 365 admin center
When
This quarter
Action
Limit app creation spaces to a pilot group
Where
Microsoft 365 admin center
Baseline checklist for agents built by business users

If you only do one row today, make it the data policy. Configure a tenant-scope policy that blocks “Chat without Microsoft Entra ID authentication in Copilot Studio” first. Sharing limits only apply to agents that require authentication. With No authentication, “You can’t control which users in your organization can chat with the agent.” Once that block is in place, sharing limits can do their job, and every other row on this list builds on it.

Get a second pair of eyes on your tenant

If you have ticked the control map against your tenant and found gaps, the next question is which ones matter for what your makers are actually building. Citizen Development Academy runs a short review. We look at what your business users have built or plan to build with AI in Copilot Studio, name the top three risks, and recommend one next step.

Book a free 30-minute AI app review with Citizen Development Academy

Stay in the loop

Get new posts delivered to your inbox. No spam, unsubscribe anytime.

Related articles