Revoke AI agent access in the Admin Console

To revoke AI agent access, use the AI agents group in the Admin Console: suspend or permanently revoke one agent, revoke a person’s delegation to it, or engage a kill switch to stop every agent of one type. The group has six pages. This guide describes each page, every button on it and every confirmation dialog. For first-time setup, read Turn on AI agents and register one.

The group appears always, but only AI overview is listed while the organization setting Non-human identities (AI agents) is off. Identities, Delegations, Approvals, Kill switches and MCP servers appear once it is on.

These pages may send you through a fresh sign-in before they show data; sign in again and you return to the page.

Lists on these pages share a few behaviors: a Search … box filters the rows already loaded, column headings that can sort do so when selected, and selecting a row opens it. When a list cannot be read, the page shows a message (for example “The delegations could not be read.”) with a Retry button instead of an empty list, so an empty list always means “none”.

Who may revoke and manage AI agents

Who	May do
Super-administrator	Everything on these pages. Only a super-administrator turns the organization setting on or off, and engages or releases kill switches.
An owner of an agent	Lists and opens the agents they own, and can suspend them. Nothing else.

The server refuses what a role may not do, and a person who may not read the organization setting sees “Only an AI agents administrator can read the setting Non-human identities, and only a super-administrator can change it.” in its place.

Overview page

The AI agents overview with agents registered

Heading AI agents, description “Register AI agents, see what they can reach, and stop them.”

Element	What it does
Register an AI agent	Opens the seven-step wizard.
Needs attention banner	Appears when something qualifies, for example “Needs attention: 1 identity has no owner; 1 identity expires within 14 days.” Links Show identities with no owner, Show identities expiring soon and Show unused identities open the Identities list filtered. “Unused” means active and not used for 30 days (an agent never used counts from when it was registered). Revoked and expired identities never count.
Card By status	Counts for Active, Pending activation, Suspended, Expired and Revoked.
Card By type	Counts for AI agent, MCP client, Automation and Service.
Recent activity	The last five AI agent events from the audit trail, each with result and time, such as “Ai agent token issued, success, …”. Empty text: “No AI agent activity yet.”
Open the audit log	Goes to Monitoring, Audit.
Organization setting, switch Non-human identities (AI agents)	Turns the feature on or off; opens the dialog below. Shows “Last changed ” once changed.

With no agents the page says No AI agents yet: “AI agents and services that call your apps.”

Dialog: Turn non-human identities on or off

Button	Effect
Type the organization’s name, then Turn on	Registered agents can get tokens again as their status, grants and kill switches allow.
Type the organization’s name, then Turn off (red)	Deletes nothing. Stops all token issuing at once; tokens already issued work until they expire, within 4.5 minutes. Engage a kill switch first to stop agents immediately.
Cancel	Closes the dialog.

Identities page and the list of agents

Heading Non-human identities: “AI agents and services that call your apps.”

Element	What it does
Search non-human identities	Filters by name, type or primary owner.
Filter by type	All types, AI agent, MCP client, Automation, Service.
Filter by status	All statuses, Pending activation, Active, Suspended, Revoked, Expired.
Clear filters	Appears while a filter or a “needs attention” link is active.
Register non-human identity	Opens the wizard. Also the button on the empty page No non-human identities yet.
Kill switches	Opens the Kill switches page.
Columns	Name, Type, Primary owner, Modes (“Acts on its own”, “Acts for a person”), Resources (a count), Last used (“Never” or a date), Expires (a date, or “Review ” when only a review date is set, or “None”), Status.

With a filter and no match: “No non-human identities match.” The needs-attention banner also appears here.

Status words

Status	Meaning
Pending activation	Saved, not yet activated. Authenticates to nothing.
Active	Can get tokens.
Suspended	Reason shown: “Suspended by an administrator”, “Suspended: its owner left. Add an owner before reactivating.”, “Suspended: its review date passed”, “Suspended by the kill switch for this type”, or “Suspended: a credential was reported compromised”.
Revoked	Permanent.
Expired	Past its Expires on date. Permanent.

An agent’s page

Select a row to open the agent. The header has a Non-human identities link back to the list, the agent’s name, its type, a status badge, and these buttons depending on status:

Button	Where	What it does
Activate	Pending	Activates it (refusals in the registration guide).
Test access	Active	Opens the Access tab and runs the test.
Suspend (red)	Active	Opens the suspend dialog. Always one click away.
Reactivate	Suspended	Returns it to Active. An agent suspended because its owner left needs a new owner first (“A non-human identity needs at least one owner.”), and one suspended by its type’s kill switch waits until the switch is released.
Menu, Revoke permanently (red)	Any state except Revoked and Expired	Opens the revoke dialog.

Tab: Overview

Identifier, Description (shown as “What the owner says it does: …” or “None”), Primary owner, Owners (“No owner” when none), Acts, Expires, Review by. Read-only.

Tab: Access

A table Resources and scopes with Application, Resource, Scopes, Until (“No end” when open-ended), or No access granted: “This identity can reach nothing yet.” Below it, the Test access section. Grants cannot be edited here; they are set when registering.

Tab: Keys

A table Public keys with Key id, Algorithm, Thumbprint and Added. A public MCP client shows “A public MCP client has no keys. It acts only for a person who has allowed it.” With two or more keys a warning explains that clients that don’t name their key in assertions are refused until only one key remains. Read-only.

Tab: Delegations

A table Delegations with Person (the user’s identifier), Resources, Scopes, Expires, Last used (“Never” until used), Status (Active, Revoked, Expired) and, for Active rows, a link-button Revoke delegation for . With none: Nobody has allowed it to act for them.

Tab: Activity

A table Activity of this identity with When, Event, Result and Done by (“Not recorded” when no administrator did it). With none: “No activity has been recorded for this identity yet.” If it cannot be read: “The activity could not be read.”

Tab: History

Registered, Last used and Status with its reason.

Dialog: Suspend

Title “Suspend ?”. Body: “Suspending stops new tokens at once. Tokens already issued stop working within 5 minutes.” Buttons Suspend (red) and cancel.

Dialog: Revoke permanently

Title “Revoke permanently?”. Body: “It can never be used again. Its keys, delegations and refresh tokens stop working. Type its identifier to confirm.” Type the identifier (for example invoice-assistant) and select Revoke permanently (red). The page then shows Revoked; the tabs stay readable and no buttons remain.

Dialog: Revoke delegation

Title “Revoke this delegation?”. Body: “This identity stops acting for that person at once.” Buttons Revoke delegation (red) and cancel. The row then shows Revoked; the agent itself stays Active.

Delegations page

Heading Delegations: “What people have allowed AI agents to do on their behalf.” A note above the list says “Every AI agent of the organization, the 500 most recent delegations. Open a row to revoke a delegation on the identity.”

Element	What it does
Search delegations	Filters by agent, person, applications or scopes.
Columns	AI agent, Person, Applications, Scopes, Expires (“Never” if none), Status (Active, Revoked, Expired).
Row	Opens that agent’s Delegations tab, where you can revoke. This page itself has no revoke button.

Empty: “No one has delegated anything to an AI agent.”

Approvals

Heading Approvals: “Scopes an AI agent asked a person to approve.” (CIBA is the name of the standard behind the per-use approval.) The page is read-only: only the person concerned decides. What the agent wrote to the person is shown only to that person, and finished requests are removed a day after they expire.

Element	What it does
Search approval requests	Filters by agent, person, scopes or resource.
Status menu	All statuses, Waiting for the person, Approved, Denied, Expired, Approved and used.
AI agent menu	All AI agents or one agent.
Columns	AI agent, Person, Scopes, Resource, Status (with the time of the decision under it once decided) and Asked.
Show older requests	Loads the next page (50 at a time).

Empty: “No AI agent has asked anyone for an approval recently.” With a filter and no match: “No approval requests match these filters.”

Kill switches page: stop every agent of one type

The Kill switches page with all four types not engaged

Heading Kill switches: “Stop every non-human identity of one type at once. Needs a recent sign-in and the type typed to confirm.” Four rows, one per type, always shown:

Column	Content
Type	AI agent, MCP client, Automation, Service.
State	Not engaged, or Engaged (with “, N suspended” after you engage it).
Since	The date engaged, or “None”.
Action button	Engage for (red) or, when engaged, Release for .

Dialog: Engage

Title “Engage the kill switch for ?”. Body: “All active identities of this type are suspended now and cannot be activated until released. Tokens already issued stop working within 5 minutes. Type to confirm.” The text you type is the type’s code: ai_agent, mcp_client, automation or service. Buttons Engage kill switch (red) and cancel.

The engage dialog asking for ai_agent to be typed

Dialog: Release

Title “Release the kill switch for ?”. Body: “Identities of this type can be activated again; suspended ones stay suspended. Type to confirm.” Button Release kill switch. After releasing, reactivate each suspended agent by hand with Reactivate.

MCP servers page

Heading MCP servers: “Applications that AI agents reach over MCP, and who may call them.” The note says they are reached over the Model Context Protocol (MCP) through an App Proxy route and are changed in the application’s Non-human callers settings.

Element	What it does
Search MCP servers	Filters by application or resource.
Columns	Application, Resource identifier, Who may call (the three Non-human callers choices), Largest request (“Default” or a number of bytes).
Row	Opens the application.

Empty: “No application is set up as an MCP server.” If some applications could not be read: “N applications could not be read, so they may be missing here.” The page reads at most 100 applications.

To add a row, set an application’s Protocol to MCP: see Connect an AI agent to Oracle E-Business Suite with the MCP Gateway.

The page also has the section MCP client registration (who may register, and the clients waiting for approval) and, on each application’s Non-human callers tab, a Token binding section. Both are described in Connect an AI agent to Oracle EBS with MCP.

The Non-human callers tab of an application

Visible only while AI agents are on. Heading Non-human callers: “Whether non-human identities may call this application, and with which scopes.”

Field	Content
Non-human callers	Refuse non-human callers, Allow, and tell the application which non-human identity is calling, Allow only on behalf of a person, and tell the application which person.
Resource identifier	An https address that names the application in tokens; must be unique. A Saved resource identifier box shows the saved value.
Protocol	HTTP or MCP. “MCP is for App Proxy routes only.”
Largest request body (bytes)	MCP only; 1 to 4194304.
Scope catalog	Per scope: Scope, Kind (Standard or Sensitive), What it allows, in words, and Tools (MCP) or Methods and Path prefixes (HTTP). Add scope (at most 64), Remove scope .
Save non-human callers	Saves the scopes, then the setting.

Choosing Refuse keeps no catalog.