
Business teams build agents in Microsoft Copilot Studio without a repository, a pull request or a deploy. That is the point of the product. It is also why most security teams cannot answer the first question AI governance asks: what is in the inventory?
Today Kodem connects to Copilot Studio, in beta. One connection inventories the agents in your tenant and checks each planned tool call against your AI policies before it runs.
No-code agents skip the controls you already own
Application security assumes a pipeline. Code lands in a repository. A build produces an image. The image runs on infrastructure you can instrument. Every step is a place to look.
A Copilot Studio agent skips all of them. A maker attaches knowledge sources, adds tools and connectors, and publishes. The agent runs in Microsoft's cloud. There is no repository to scan and no host to put a sensor on.
So the inventory comes from a survey, or from whoever remembers. And a list of agent names is not an inventory. It does not say which connectors an agent holds, which tools it can call, or what it is wired to.
The evidence already exists. It sits in the platform.
Kodem's rule is that evidence beats declarations. For code you host, that evidence is what is loaded in the running process, which is what runtime intelligence reads. On a platform you do not host, there is no process to read. The evidence lives in two places: the platform's own record of each agent, and the moment an agent is about to call a tool.
The Copilot Studio connection reads the first and sits at the second.
One connection inventories your agents
Kodem signs in to your Microsoft tenant as an Entra app registration that you create, with a federated credential. You give Kodem no secret. The app needs read access to the Dataverse tables that describe your agents. There is no sensor to deploy.
From there Kodem lists the agents in your Power Platform environments, published or not, each with its topics, knowledge sources, tools, MCP tools and connectors. It checks again every five minutes. Grant read access to conversation transcripts, which is optional, and Kodem also shows each agent's last activity.
These agents land in the same AI Discovery inventory as the agents Kodem finds in your code, your images and your running workloads. Kodem also checks each agent's configuration for risk, such as an agent published without authentication, or tools that run on the maker's credentials.
Each tool call is checked before it runs
Inventory shows what an agent can do. Policy decides what it may do.
Copilot Studio lets you bring an external security provider. Turn it on, and Copilot Studio asks Kodem before an agent runs a tool, a connector or an MCP call. Kodem evaluates the call against your AI policies and answers allow or block. A blocked call does not run, and Copilot Studio tells the user.
Policies match on tool name, tool category, connector and tool arguments. They can also act when prompt injection, personal data or secrets are detected. Each policy runs in block or monitor mode. Monitor records the match and lets the call through, which is where most teams should start.
Every decision lands in the violation log, tied to the policy, the agent and the tool behind it. This is a record of what happened: a real call, allowed or denied.
The boundaries are part of the design
A control that sits in front of business agents has to be plain about its limits. These are ours.
- It fails open. If Kodem does not answer within one second, or cannot be reached, Copilot Studio allows the call by default. Your agents do not stop because a security check was slow.
- It stores decisions, not conversations. Copilot Studio sends call context to its security provider so each call can be evaluated. Kodem evaluates the tool definition and its inputs, and stores only the decision: the tool and connector, the policy, and whether the call was allowed or blocked.
- The decision point is the tool call. Kodem does not run inside Microsoft's cloud. The check applies to agents that use generative orchestration, which is where Copilot Studio calls its security provider. For custom tools and MCP servers that run on your own compute, the Kodem runtime sensor sees what each tool does.
- A policy is a control, not a promise. You write the policies, scope them and turn them on. New policies start disabled. A policy blocks what you told it to block.
- It is a beta. The connection is open to tenants in the AI Discovery beta, on Kodem SaaS. Kodem reads environments that have Dataverse, and each Microsoft tenant connects to one Kodem tenant. Microsoft offers its external threat detection interface in preview.
A no-code agent is still an application
A Copilot Studio agent holds credentials, reaches data and takes actions. That makes it an application, whoever built it and wherever it runs. It belongs in the same inventory and under the same policies as everything else you operate. That is the argument behind agentic AI security. It is also why runtime AI security and runtime application protection are converging into one discipline rather than two security stacks.
Kodem does not run your governance workflow. It supplies the evidence and the decision that feed it: which agents exist, what each one is wired to, and which calls were denied.
The setup steps are in the Copilot Studio connection guide. To join the beta, talk to your Kodem team.
Frequently Asked Questions
Through a cloud connection with read access. Kodem signs in to your Microsoft tenant as an Entra app registration that you create and reads the Dataverse tables that describe your agents. It lists each agent with its topics, knowledge sources, tools, MCP tools and connectors, and checks again every five minutes. No sensor is deployed.
No secret. The app registration uses federated credentials, so Kodem stores no Azure secret. It needs read access to the Dataverse tables for agents, their components and their connection references, and Kodem recommends a custom read-only role. Disconnecting stops discovery and tool-call checks, and Kodem keeps nothing in your tenant.
Yes, per tool call. With Copilot Studio's external threat detection turned on, Copilot Studio asks Kodem before an agent runs a tool, connector or MCP call, and Kodem answers allow or block by your AI policies. A block denies that one call. It does not stop or delete the agent. A policy can also run in monitor mode, which records the match and allows the call.
The check fails open. If Kodem does not answer within one second, or cannot be reached, Copilot Studio allows the tool call by default. The control is built so that a failure of the security check does not stop your agents.
No. Copilot Studio sends call context to its security provider so each call can be evaluated. Kodem evaluates the tool definition and its inputs and stores only the decision: the tool and connector, the policy, and whether the call was allowed or blocked. Inputs and conversations are not stored.
It is in beta, for tenants in Kodem's AI Discovery beta, on Kodem SaaS. Kodem reads Power Platform environments that have Dataverse. Microsoft offers its external threat detection interface in preview.
Related blogs

When Everything Becomes Open-Source
Our research shows that an AI agent can reconstruct a working copy of an application when given the original software's binary or package.
13
Stop the waste.
Protect your environment with Kodem.
The State of the Application Security Workflow
This report aims to equip readers with actionable insights that can help future-proof their security programs. Kodem, the publisher of this report, purpose built a platform that bridges these gaps by unifying shift-left strategies with runtime monitoring and protection.
.avif)
Get real-time insights across the full stack…code, containers, OS, and memory
Watch how Kodem’s runtime security platform detects and blocks attacks before they cause damage. No guesswork. Just precise, automated protection.



