On October 19, Microsoft published a post arguing that organizations should secure AI agents the way they secure insiders: identities with legitimate access that can still cause harm. I think the framing is right. It's also more useful than most agent security writing, because it starts from the assumption that authentication succeeded and asks what happens next.
An agent in a Microsoft 365 tenant is authenticated, authorized, and acting with real permissions. Many of the ways it goes wrong involve no break-in at all. Microsoft's post highlights two: an agent that reward-hacks its way to an objective, and an agent that lacks the organizational judgment a human employee would apply, with no compromise involved. Redirection through malicious retrieved content, known as indirect prompt injection, is a separate failure mode that Microsoft's Zero Trust guidance addresses.
Below: what Microsoft proposed, where documented Microsoft 365 controls apply today and where they stop, and a checklist that doesn't depend on any announced product shipping.
A valid sign-in was never proof of good behavior
Insider risk programs never treated a valid sign-in as proof of safe behavior. A finance analyst with legitimate SharePoint access can still download the wrong folder or share it with the wrong person. The program asks who this is, who vouches for them, what they can reach, what they normally do, and how fast you can pull their access.
Swap "analyst" for "agent" and every one of those questions still works. A successful token request tells you the identity met the conditions for a token. It tells you nothing about whether a specific action is appropriate. Agents widen that gap because they assemble their instructions at run time from several sources: the system prompt, the user's request, retrieved documents, and tool output.
These failure modes need different responses:
- Compromise through content. Indirect prompt injection plants instructions in a document, email, or web page the agent retrieves. Success depends on the model, the orchestration, the agent's permissions, and the defenses in place. It doesn't work every time, and an attacker only needs it to work occasionally.
- Reward hacking, where the agent finds a path that satisfies the measured objective without doing the intended work. No attacker required.
- Gaps in organizational judgment. An employee knows a draft board deck shouldn't go to an external partner even when sharing is technically allowed. An agent knows what its instructions and permissions tell it, and not much else.
The second and third modes are why the insider framing beats a pure prompt-injection framing. Better input filtering does nothing for an agent that is working as designed and still doing the wrong thing.
The judgment gap is the one that's hardest to dismiss. Nobody attacked anything, every permission check passed, and the board deck still left the building.
Automation can also compress the window for response. People can script data access too, so this isn't unique to agents. The difference is that fast, multi-step action is what an agent does by default, while a person has to choose to build it. I don't have a measured comparison to offer, and I'm not going to invent one, but quarterly access reviews were designed for a much slower actor.
Five questions for Project Perception and Guardian Agents
Microsoft's answer is Guardian Agents, which it describes as moderating every action of every agent before it happens. The idea is a supervision layer that watches what agents do while they do it, instead of relying on manual review after the fact. The direction is correct. Access reviews were designed around humans whose roles change over months, and agents don't work on that clock.
The effectiveness claims belong to Microsoft. The post describes intent and capability. The announcement's effectiveness claims need independent validation, and I'd treat them as vendor commitments until that exists.
Before planning around either piece, these are the questions I'd want answered:
- Coverage. Which agent platforms does it observe? Microsoft-hosted agents are the obvious case. Third-party agents calling Graph through their own orchestration are harder to see.
- What deployment requires: licensing, telemetry prerequisites, and whether agents must be registered with a particular identity type.
- Microsoft describes pre-execution moderation. Which actions and platforms support that enforcement, and what testing demonstrates its effectiveness?
- A guardian built on a model inherits some model weaknesses, including exposure to the content it reads. How is that handled?
- Status and timeline. Proposed, preview, or generally available, and for which tenants.
I'd ask these before counting any announced capability as a control. Until someone answers them, today's plan has to rest on controls that are documented and testable in your own tenant, starting with who each agent actually is.
Every agent needs an owner who is a person
The first insider-risk question is who. Every agent needs an identity you can trace its actions back to, and a named human owner accountable for it. That means a person. A distribution list doesn't count.
The correct access pattern depends on what the agent does. When it acts on behalf of a signed-in user, delegated permissions are the right fit: the agent works within that user's access and that user's sign-in context. When it performs autonomous work with no user present, app-only access is appropriate, scoped as narrowly as the workflow allows. Set up properly, you can attribute and revoke either pattern. The failure mode is an agent running on a shared account or a credential with no owner.
Conditional Access applies differently to each pattern, and the scopes are easy to blur:
- For Entra agent identities, Microsoft's Conditional Access for Agent ID documentation lists Microsoft Agent 365 plus at least Entra ID P1 or Microsoft 365 E3, or Microsoft 365 E7.
- Creating or modifying policies scoped to eligible service principals requires Workload Identities Premium. That requirement applies to that policy scope only. Not every agent access pattern needs it.
- Delegated access, where the agent acts for a signed-in user, falls under the user-targeted policies that apply to that user.
Conditional Access evaluates when a token is requested, including on-behalf-of exchanges. Continuous Access Evaluation lets certain events, such as revocation, take effect sooner for supported resources. Conditional Access answers whether this identity should get a token under these conditions. Whether this agent should send this file to this destination is a separate question, and other controls have to answer it.
Sites.Read.All is a contractor with every key
Insider risk programs worry most about broad standing access. Agents are easy to over-permission, because a tenant-wide scope like Sites.Read.All gets a prototype working on the first try. That's a contractor with a key to every building. If I had to bet on which failure shows up first in a real tenant, it's this one, and the model has nothing to do with it.
Sites.Selected is the better pattern for SharePoint, and it works in two steps. Microsoft's Sites.Selected overview lists three steps: consent, an administrator's explicit grant of the app to specific resources at a specific role, and the token request. It states that all three must be completed to get the expected access. Microsoft Graph's selected-permissions model also covers lists, list items, and files through their own scopes, which helps when a workflow needs less than a whole site.
Scope to the approved workflow. Observed usage shows what the agent actually touched. Foreseeable legitimate needs show what it will need at quarter-end. Strip everything unused this month and you break the month-end report. Leave everything granted at build time and you keep standing access nobody approved.
On the content side, SharePoint offers controls for limiting what users and apps can discover and who can reach a site, including Restricted Content Discovery (https://learn.microsoft.com/en-us/sharepoint/restricted-content-discovery) and Restricted Access Control (https://learn.microsoft.com/en-us/sharepoint/restricted-access-control). Restricted Content Discovery limits whether a site's content appears in organization-wide search and Copilot experiences. It does not change existing permissions; users who already have access can still open the content directly. Check availability and licensing for your tenant before building a plan around them; I wouldn't assume either is present.
One planted instruction, five places to stop it
This scenario is hypothetical. Picture an agent that summarizes documents for a project team, with permission to read one project site and send email. One document contains planted text instructing the agent to email a summary of the folder to an external address.
Whether anything happens depends on a chain of conditions. The model has to follow the planted instruction. The orchestration has to fail to separate retrieved content from instructions. The agent's mail permission has to allow external recipients. Transport rules or data loss prevention have to miss the outbound message. A configured blocking control would need to fail to stop the send. An alert alone may arrive without preventing it.
Mail.Send made the send possible. It didn't authorize sending confidential material to an outsider, and the audit trail will show an agent acting outside its purpose. The controls between possible and happened are the actual subject of agent security.
Microsoft's Zero Trust guidance on indirect prompt injection treats model-level defenses as one layer among several. That's the right posture. Model defenses reduce the odds; deterministic controls enforced outside the model decide what's reachable when they fail.
Search the audit log before an investigation does
Insider risk programs run on records. Microsoft's documentation says Purview auditing is on by default for Microsoft 365 organizations, with explicit exceptions: tenants on Business Basic, Business Standard, and Business Premium licenses, and unmanaged enterprise trial tenants. If your tenant falls into one of those groups, don't assume auditing is on. Check.
Confirming the toggle is the easy part. The harder checks:
- Retention for the audit tier you're licensed for, measured against how far back an investigation might need to look.
- Run the agent, then search the audit log for its identity. Confirm the log actually records the events you'd need for an investigation.
- Name the person who owns alerts on agent behavior. An alert landing in an unread shared mailbox is functionally no alert.
- Confirm the relevant protection is configured for blocking rather than audit-only operation, and that it covers the agent's actual tool path.
For runtime protection, Microsoft documents threat detection and protection for AI agents in Defender XDR, including detections for indirect prompt injection and suspicious agent behavior. Microsoft's Defender documentation distinguishes a default audit rule, which records activity without stopping it, from custom blocking rules. Detection alone is not an enforcement barrier. Coverage depends on the agent platform, available telemetry, supported platforms and tools, integrations, and configuration. The documentation lists prerequisites and marks which capabilities are in preview. Read it against the specific agents you run instead of assuming they're covered.
Approval gates the agent can't approve
For high-risk actions, external data transfers and permission changes above all, require an approval enforced outside the model. A few properties matter. The agent can't approve its own request. The agent can't reach the same outcome through a different tool. The approver sees what is actually about to happen.
Adding a Power Automate approval or a Teams adaptive card doesn't guarantee any of that. If the agent's identity can complete the action directly, the approval is advisory. If the approver only sees a summary the agent wrote, the approval inherits the agent's framing. The gate has to sit where the action executes, and the agent's identity must lack the permission to skip it. Then try to skip it anyway.
What I'd do this month, in order
Nothing here waits on Project Perception or Guardian Agents. This is the order I'd work through it:
- Inventory every agent. Start with the Agent 365 inventory and Entra identity views, then reconcile those with Copilot Studio, enterprise applications, and the other agent platforms in use. Reviewing app registrations with Graph permissions can identify additional integrations, but it doesn't prove every agent has been found.
- Each one gets a named human owner. Disable anything nobody will own.
- Confirm the access pattern per agent: delegated when it acts for a signed-in user, narrowly scoped app-only when the work is autonomous.
- Move SharePoint access from tenant-wide scopes to Sites.Selected with explicit grants wherever the workflow allows.
- Map which Conditional Access scope applies to each agent, and confirm your licensing actually covers that scope.
- Verify audit status, retention, agent event coverage, and alert ownership in your tenant.
- Check that the relevant protection is set to block rather than audit-only, and that it covers the agent's actual tool path.
- Put approval gates outside the model on external transfers and permission changes, then attempt to bypass them.
- Run controlled prompt-injection tests in a non-production tenant, with planted content in documents and email the agent reads. Record which layer stopped each attempt, or that none did.
Agents belong in the insider risk program
Microsoft is right that agents belong inside the insider risk program. The framing holds because it doesn't depend on catching an attacker; reward hacking and judgment gaps happen without one. Continuous supervision through Project Perception and Guardian Agents is a reasonable direction. Whether it delivers what the post describes is something I'll judge from documentation and independent testing.
My bet is that runtime supervision earns its place on the judgment gaps, the board-deck cases no permission model can express, and becomes one more ignored queue in any tenant that never named an owner for its alerts. That's why I'd start with the boring layers. Attributable identity, scoped access, verified audit, and gates outside the model are things an admin can configure and test this month, and I'd want them in place whatever ships next.
If you've run the inventory in your own tenant: how many agents were still on Sites.Read.All or another tenant-wide scope, and how many of those had an owner who knew it? And when you tried to route around an approval gate, did the agent's identity let you through?