Agent Perimeter

Start here

What this tool does, what makes its numbers different, and the order to read it in.

Agent Perimeter finds the AI agents in a Microsoft 365 tenant — Copilot agents, service principals, anything with an identity that can act — and tells you which of them nobody owns and nothing is stopping.

It reads. It does not change anything in your tenant: there is no apply path in this panel and there is not meant to be. Remediation is a separate script you run deliberately.

Three things to know before you read a number

It counts agents Microsoft’s own enumeration misses

The documented agentIdentity type cast is what Microsoft-native inventories list. Agents that predate it, or were never converted, are not returned by it — they are not hidden, they are simply not counted. On the demo estate that is 393 of 551. Expect this tool’s total to be larger than the one your other tools show, and expect that difference to be the interesting part.

“Not checked” is an answer, and it is not a pass

“No agent is risky” and “the risk read was refused because this tenant has no Entra ID P2” produce the same zero. Every check here declares what it needs, and reports Not checked when it did not get it. Those are not near-misses and they are not clean — they are the parts of the estate nothing has looked at.

Observability is not prevention

Most coverage figures count an agent as covered when it merely appears in a log. That tells you afterwards; it stops nothing. This tool separates the two, which is why an estate can read as 97% covered here and still have zero agents that anything would actually stop.

The order to read it in

Run an assessment

Read-only, against the tenant you are signed in to. Takes a few minutes. Everything else on this panel is a view of what it found.

See what is wrong, and the route out

The Overview carries the headline finding and an ordered list of what closes the most checks. Posture is the same findings as named checks, each one linked to the admin blade that fixes it.

Check what the run could NOT read

Do this before you quote any figure. Evidence lists every source the run touched, every one it was refused, and what each gap does not prove. A number derived from a source that was not read is not a smaller number — it is no number.

Drill into the agents themselves

Search, sort and filter the estate, and open any agent to see what it can reach, who owns it, and why it scores what it scores.

Record what you are going to do

Actions tracks the decision and the owner. It deliberately does not perform the work.

What it will not tell you

Whether an agent is behaving badly. It reports what an agent is permitted and exposed to, plus whatever signals the tenant actually records.
Anything about egress unless you have told it what your network controls are. Restricted egress is reported as unmeasured, not as absent.
Anything it was refused. Every refusal is listed on Evidence rather than smoothed into a lower number.

Overview

What is in the estate, what holds it, and the shortest route to a clean result.

loading…

Run an assessment read-only against the tenant

running…

NextRead the findings as named checksEverything above, one row per check, each with the severity, the agents it affects and a link to the admin blade that fixes it.Open →

Start here

Three things to know before you read a number

The order to read it in

What it will not tell you

Overview

Agents

Agent posture

Threat intelligence

Control coverage

Control coverage, per agent

Actions

Evidence

What was read, and what was not

Run history