When an agent misbehaves, a tool is compromised or a policy turns out to be wrong, you need to stop calls quickly and without breaking what you can still trust. K-Block Cloud has several controls for that. They differ in scope (one call, one agent, one tool, every agent of a policy), in who may use them, and in how they can be undone.
This page lists every operational control that exists in this version and helps you pick the narrowest one that solves the problem. Their use is recorded in the audit log, with one gap: Close the remaining open actions records the actions it closes, but not who used it.
Which control to use
| Situation | Use | Scope | Who |
|---|---|---|---|
| One pending or approved call must never run | Cancel the action | one action | owner, admin |
| A policy lets through what it should not, for many agents | Emergency stop | every agent of one policy | owner, admin |
| A policy change was a mistake | Roll back | every agent of one policy | owner, admin |
| Observe mode relaxes decisions you no longer want relaxed | End observe windows | one agent, or all | owner, admin, developer |
| One agent behaves badly | Disable the agent | one agent | owner, admin, developer |
| One agent's key leaked | Revoke the API key | one key | owner, admin, developer (gateway keys) |
| One tool or MCP server must not be called | Disable the tool or server | one tool, or every tool of a server | owner, admin, developer |
| An agent must lose its authority for good | Revoke its delegation | one delegation | owner, admin |
Cancel an action
A cancellation ends one protected action that is pending approval or approved but not yet executed. It is a CANCELLATION signed by your organization's principal, so the action can never be approved or executed afterwards, and K-Block issues a receipt for it.
- Open the action under Actions.
- In Cancel this action, optionally enter a Reason (optional); it is signed into the cancellation.
- Select Cancel action…, then Sign cancellation.
You need the organization principal's key in this browser; without it the card says so and links to Organization settings. Only owners and admins see the card. An action that is already executing cannot be cancelled, and a cancellation that loses a race with an execution or a decision is refused with the action's current state. See Actions.
Emergency stop
An emergency stop replaces a policy's active version with one that denies every call (Deny all, recommended) or requires approval for every call (Require approval for all). It takes effect on the next call on every Gateway.
On the policy's page, select Emergency stop…, choose the mode, decide what happens to open actions (calls waiting for approval, and calls approved or allowed that have not run yet), type the policy name and select Deny every call now (or Require approval for every call now). The stop:
- creates and activates the new version at once; the previous active version is superseded;
- ends the observe windows of every agent that uses the policy, and no observe window can relax an emergency version;
- re-checks calls already allowed without approval before they run, so a stopped policy lets nothing through outright;
- with Close pending approvals and approved calls that have not run (default for deny all), closes the open actions of the replaced versions. With Close pending approvals only or Keep them (default for require approval), calls a person approved before the stop can still run (with Keep them, pending requests can also still be approved): cancel the ones that must not run.
The stop covers one policy and the agents assigned to it, not the whole organization. Undo it with a rollback once the incident is resolved. Full details are in Emergency stop.
Roll back a policy
A rollback re-activates an earlier version of a policy (by default the one active before the current one) as a new version. It takes effect on the next call. By default open actions are kept. In the dashboard you can also close the approvals still pending; through the Cloud API (open_actions: "supersede_all") you can also close the calls a person approved that have not run.
A rollback does not end observe windows. If closing open actions stops before all are closed, the result offers Close the remaining open actions; K-Block also closes the rest automatically, checking for them every 15 seconds. See Activate, roll out and roll back.
End observe windows
An observe window lets one agent's calls through where its policy would deny them or ask for approval, for a limited time (see Observe and shadow). Ending a window restores normal enforcement at once.
- One agent: select End now on the agent's page or in the observe banner at the top of the dashboard. Owners, admins and developers can do this, and so can a management API key with the
agents:writescope. - Every window of the organization: call the Cloud API, for example with a management API key that has the
agents:writescope. The dashboard has no button for this.
POST /api/v1/orgs/<org-id>/observe/end-all
Authorization: Bearer kb_live_…The answer lists the windows it ended ({ "ended": [{ "window_id": …, "agent_id": … }] }). A signed-in owner, admin or developer may call it too.
The owner has a stronger switch in Settings under Observe mode: setting Longest observe window (hours) to 0, or clearing Allow observe windows (the live opt-in), ends every open window and prevents new ones; lowering the maximum ends the windows that are longer than the new limit. Windows also end when an emergency stop hits their policy, when their agent is disabled or switched to another policy, and at their time limit.
Disable an agent
On the agent's page, set Status to Disabled (every call refused) and save. From the next request, the Gateway refuses every call with any of the agent's gateway keys as unauthenticated (HTTP 401), including the retry of an approved call, so nothing is consumed. Disabling also ends the agent's observe window.
Its open actions stay as they are: they cannot execute while the agent is disabled, but the agent can still execute them if you enable it again before they expire. Cancel the ones that must never run. Owners, admins and developers can disable an agent, and so can a management key with agents:write. See Agents.
Revoke an API key
Under API keys, select Revoke on the key and confirm ("Requests with it fail immediately."). The Gateway authenticates every request anew, so the next request with a revoked gateway key fails with HTTP 401, even on a connection that is already open.
Owners, admins and developers can revoke gateway keys; only owners and admins can revoke management keys. Revocation cannot be undone: create a new key for the agent. K-Block also revokes keys automatically when the member who created them is removed, leaves or loses the role needed for that kind of key. See API keys.
Disable a tool or an MCP server
- A tool: edit it under Tools & MCP servers and clear Enabled (callable through the gateway).
- Every tool of a server: on the server's page, set Status to Disabled (its tools are not callable).
A disabled tool disappears from the agent's tool list, and calls to it are refused as an unknown tool (UNKNOWN_TOOL) without being recorded. An approved call of a tool that was disabled since is refused the same way when the agent retries it, so nothing is consumed. Owners, admins and developers can disable, and so can a management key with tools:write; enabling a tool again requires a signed-in owner or admin, because it authorizes the tool's binding. Setting a server back to Active does not need one. See Tools and MCP servers.
Revoke a delegation
A delegation is the principal's signed grant that lets the Gateway act for an agent (see Delegations). Revoking it removes the authority it granted: pending and approved actions under it can no longer execute, and unless another active delegation of the agent grants the same actions, the tools it covered are no longer listed and new calls are denied with NO_DELEGATION.
On the agent's page, select Revoke on the delegation, then Sign revocation. The revocation is signed by the organization principal in your browser, so owners and admins with the principal's key can do it.
If a call under the delegation is running at that moment, the delegation is withdrawn at once (no new call can use it and no approved call under it can execute), but the revocation itself is refused. Once the running call has finished, select Sign revocation again. A revocation cannot be undone: sign a new delegation to restore the agent's authority.
After an incident
- Check Actions for calls that were allowed or approved before you acted, and Receipts for what executed.
- Read the audit log to see who changed what. Actions closed by a stop or rollback appear as
action.supersededwith thesystemactor. - Undo the temporary controls: roll back the emergency stop, enable the agent or tool, start a new observe window if you still need one.
- Replace anything that leaked: rotate the upstream credential, create new API keys, rotate webhook secrets.