We spent years building insider-risk programs for employees who might go rogue: access reviews, joiner-mover-leaver processes, and someone watching for the payroll clerk who suddenly downloads the whole HR share. This week, the company selling a lot of us our AI agents said that playbook now applies to software too. That software doesn’t sleep, doesn’t take PTO, and may still hold the admin rights it got in a pilot nobody went back to. Insider threat security for AI agents is no longer a thought experiment. It is a vendor talking point. Before I turn it into a to-do list, it is worth reading what Microsoft actually said.
AI Agents and Insider Threat Security: What Microsoft Actually Said
On October 19, 2026, Microsoft published two posts. On the Official Microsoft Blog, Hayete Gallot, Executive Vice President of Microsoft Security, wrote “Securing the agentic insider.” She writes that “frontier models, regardless of their origin, cannot be trusted,” and that “This is not because models are inherently malicious.” Her argument is that models “have been trained to attain results, which often leads to reward hacking.” She says agents “are frequently overprivileged, have access to data, tools and systems and can take actions and make decisions in critical processes.” Then comes the line the coverage picked up: “They are agentic insiders.”
The same day, Nadim Abdo, Corporate Vice President for Identity and Network Access, published “Agents are your newest insiders. Security must move at their speed.” on the Microsoft Security Blog. His wording is more measured: “With that capability can come insider risk. This does not mean agents are inherently malicious. It means they are capable tools that can be granted access to important systems by users. They can make mistakes, be manipulated, or be compromised.”
Notice what is missing. Neither post, as I read it, says “insider threat.” Microsoft says “insiders,” “agentic insiders,” and “insider risk,” and both executives say outright that this is not about malice. The “insider threats” framing is TechCrunch’s paraphrase. TechCrunch writes that Satya Nadella has been warning the risk may be agents enterprises deploy, “much like a rogue human employee, otherwise known as an ‘insider threat.’” That is a fair summary of the direction, but it is not Microsoft’s wording. That is why I softened the headline to “newest insiders.”
The Principle Underneath
The core idea in Abdo’s post is one any operator will recognize: “surround non-deterministic intelligence with deterministic controls, independent evidence, and containment boundaries. Assume an agent can fail or be compromised.” He quotes Nadella: “The most trustworthy agentic system will not be the one with the model we trust most. It will be the one designed that enables us to trust the model least.”
This is the same reason we never let an application validate its own audit log. The thing being watched cannot be the thing doing the watching. Abdo puts it as “the model cannot be its own security boundary.”
What Microsoft Has Documented vs. What It Just Announced
This is where the Microsoft announcements get blurry, so I am splitting them into two groups.
Documented Today
- Microsoft Entra Agent ID. Microsoft’s Entra security for AI overview on Microsoft Learn describes agent identities, agent identity blueprints as templates, an agent registry that lets you “assign identities to agents, autodiscover them across your organization, and manage all agent metadata in one place,” and logging: “All authentication and actions performed by agents are logged in Microsoft Entra ID.”
- Conditional Access and ID Protection for agents. The same page says Entra applies Conditional Access to agents, including “Microsoft Managed Policies providing a secure baseline by blocking high-risk agents,” and that ID Protection flags “anomalous activities involving agents.”
Neither Learn page I read carries a preview label on these features. Still, check your tenant and licensing before you plan around them.
Announced October 19, Availability Not Stated
Abdo’s post says Microsoft is “introducing” the following. It gives no general-availability dates, so I am treating them as announcements, not shipped products:
- Dynamic Permissions: task-scoped, time-bound authority, where “Higher-impact actions can require verified human approval.”
- Dynamic Agent Risk and Dynamic Conditional Access for agents. Abdo describes the response as “graduated, not binary,” for example preserving reads while blocking writes.
- Guardian Agents and Playbooks in Project Perception.
- Wider discovery: 35 more local agent types, for “60 supported local agents, including Claude Code, OpenAI Codex, and GitHub Copilot.”
- New Entra Agent ID identity types for local agents found through sensors. This is an extension of the existing product.
The post also mentions Microsoft Execution Container (MXC) as a sandbox example, again without availability details. These claims about Microsoft’s own products are Microsoft’s, so treat them as vendor claims.
My Recommendations: An Identity, a Badge, and a Manager
This section is my own recommendation, not Microsoft’s guidance, and it applies whatever vendor you use.
- Give every agent its own identity. Abdo writes that an agent “should not inherit every permission of the employee who sponsors it.” If your agent signs in as a person or a shared service account, you cannot tell who did what. Fix that first.
- Make the badge narrow and short-lived. Start with read-only access to the specific systems the task needs. Add write access deliberately. Put an expiration date on anything granted for a pilot.
- Assign a human manager. Every agent needs a named owner who approves its access, answers for its behavior, and gets the alert when it misbehaves. Unowned agents are the “agent sprawl” Microsoft’s Learn page warns about, including agents “created for temporary purposes” that “remain in production indefinitely.”
- Run access reviews for agents. Put them in the same quarterly review as people. If nobody can explain why an agent has a permission, remove it.
- Write the offboarding runbook now. Know how to disable an agent’s identity, revoke its tokens, and find what it touched before you need to.
- Keep monitoring outside the agent. Logs and kill switches should live where the agent cannot change them.
Here is a hypothetical to make it concrete. A summer pilot gives a reporting agent a global-admin-equivalent role “just to get it working.” The pilot ends, the intern who built it leaves, and the agent keeps running nightly with those rights. Nothing malicious has happened, and nothing has to for that to be a problem. It is my illustration, not a reported case.
The Bottom Line
Microsoft did not say your agents are evil. It said they are insiders: capable, often overprivileged, and fallible. I think that framing is right. It is also, as TechCrunch notes, a framing that helps sell Microsoft’s security products. You don’t need to buy anything to act on it, though. Give each agent an identity, a narrow badge, and a manager, and review all three on a schedule.

Leave a Reply