Agent Insider: Microsoft details how to protect AI agents within companies

Microsoft published an analysis of a risk that companies began to map: the “agentic insider”, the AI agent that operates within the organization with its own identity, real permissions and access to corporate data. The central conclusion is that these agents need to be treated as digital employees — with credentials, limited scope and audit — and not as any software resource. The analysis runs through Copilot, Power Platform and Azure, precisely the platforms where these agents are already operating.

What is the agency insider?

The term displaces the classic “insider Threat” — the official who causes damage by error or bad faith — into the terrain of autonomous agents. The difference is that an agent does not act willingly: he executes what was set up, within what was authorized. The problem arises when this authorization is too generous, as Microsoft details in the original post.

In practice, an agent connected to corporate email, SharePoint and a call system can read, summarize, forward and fire actions in sequence without anyone following each step. If he inherits the permissions of a human user, he also inherits everything that this user sees — including what he should not. Add to this vectors already documented by Microsoft itself, as malicious instructions hidden in pages and documents, recurring theme in research on AI agents with web access.

How Copilot, Power Platform and Azure enter the agency insider

The content deals with the three layers where the problem materializes. In Copilot, the agent responds based on what the user can already access — which makes permission control the first firewall. At Power Platform, agents built internally can gain their own identities and act on business systems. At Azure, stay the accommodation, secrets and connections with the rest of the infrastructure.

The recommendation that crosses the three: explicit identity for each agent, minimum permission required, conditional access by context and complete record of what he did. It is identity governance applied to a new type of user who does not quit, but multiplies quickly.

Why the agency insider has now become a priority

Agents have left the demonstration stage. With them, the number of non-human identities within a company grows faster than the ability to manually review them. Each new agent is a new credential, with its own scope — and if no one would invent it, no one would revoke it when the project is over.

Microsoft was preparing the ground: it published Zero Trust’s tools specifically for AI agents and treated trust as a pillar of AI adoption. The analysis of the agency insider is the next step, by turning the discourse into an operational checklist for those who manage corporate environments.

What the agency insider changes to Brazilian companies

For those operating under LGPD, the agent is another point of processing personal data — and must enter the same inventory used for traditional systems. In practice: list each active agent, map which bases it reaches, define who answers for it and review permissions periodically. Regulated sectors, such as banks and health, tend to be charged first, because they already meet formal traceability requirements.

Adopting this early costs less than fixing it later: agent without defined owner is agent that nobody shuts down when the budget changes.

What Agentic Insider Analysis Doesn't Solve

It is worth the message about the scope: it is not a new product, nor a benchmark. Publication is guidance — it does not bring deadlines, prices or effectiveness metrics. Nor is there a specific norm for autonomous agents in Brazil; the adequacy remains depending on the configuration of each client and the seriousness with which the teams apply the suggested controls.

Frequently Asked Questions about agentic insider

What is an “agentic insider”?

It is an AI agent that operates within the company with its own identity and permissions, able to access data and perform actions — the digital equivalent of an internal user.

Microsoft launched a new product?

Nope. Publication is a security analysis with identity, permission and governance recommendations for agents running Copilot, Power Platform and Azure.

Does my company have to act now?

If there are already agents in production with access to internal data, the minimum is to inventory these agents, limit permissions to what is needed and record everything they do.