Quick answer: Revoking Gazebo access stops future credential retrieval. Revoking or replacing a provider key addresses copies already delivered. Choose the action—or combination of actions—based on what may be compromised, rather than treating rotation as a failure to set up revocation.
Three operations that are often called “revocation”
| Operation | What changes | What it does not establish |
|---|---|---|
| Retrieval-access revocation | Gazebo rejects future retrieval using the revoked agent access or connection | An already-delivered provider key is not invalidated |
| Provider-key revocation | The provider stops accepting the credential, subject to its documented behaviour | Completed operations are not undone; other access paths may remain |
| Provider-key rotation or replacement | Consumers move to a replacement key and the old one is retired | Creating the replacement alone does not disable the old key |
Always identify which credential is being revoked. A Gazebo agent token and the Stripe key returned after authentication have different issuers and different consequences.
When retrieval revocation is useful
An agent may be retired, a connection may no longer be needed, or you may want to prevent further key delivery while investigating an incident. Revoking that retrieval access is a targeted action: other Gazebo profiles can remain active without changing their identities.
Gazebo checks the retrieval policy and returns the configured stored provider key when authorised. The runtime and its host receive the actual value and may retain it. Revocation therefore blocks the next retrieval, not necessarily the next provider call.
That is still useful access management. It is not complete containment when a potentially compromised runtime already holds a valid key.
When to invalidate the provider credential
If a provider key appears in a public repository, an untrusted conversation, an attacker-controlled destination or a compromised runtime, assume that deleting the visible copy may not remove every copy.
Restrict, revoke or replace it at the provider as appropriate to the incident. Stop or isolate the affected runtime where possible. If the provider supports temporary credentials, check their actual expiry and revocation semantics; a short lifetime is not a reason to ignore active misuse.
Changing a Gazebo profile's declared-method policy does not narrow a write-capable key already delivered. Provider permissions determine subsequent direct API access.
Rotation does not always require an outage
Some providers support multiple active keys or an overlap period. For a routine replacement, a staged sequence may be possible:
- Create a replacement with the intended provider permissions.
- Update the stored value and known consumers through the normal deployment process.
- Verify retrieval and legitimate provider use without exposing either key in logs.
- Retire the old key and verify that it is no longer accepted.
- Update the inventory and review any consumers that were missed.
This sequence depends on provider support. During an active compromise, immediate invalidation may be more important than a graceful cutover. Do not keep an exposed key active solely to avoid disruption.
Rotation can reveal unknown dependencies. Maintaining an owner and consumer inventory makes that operational risk easier to manage; it does not eliminate it.
Response table
| Situation | Response to consider |
|---|---|
| An agent is misbehaving but no exposure has yet been established | Block retrieval, stop or isolate the runtime, and assess whether it holds a key that also needs provider invalidation |
| A provider key was exposed or the runtime holding it may be compromised | Invalidate or restrict it at the provider; revoke affected retrieval access and coordinate replacement for legitimate consumers |
| Only an agent's retrieval token was exposed | Revoke or replace that token; investigate whether it was used to retrieve provider keys and invalidate those where needed |
| An agent is being retired | Revoke retrieval, retire the runtime, and invalidate dedicated provider credentials when continued use should end |
| The task needs fewer permissions | Narrow provider credentials and retrieval policy separately; account for broader keys already delivered |
| A scheduled provider-key replacement is due | Use a tested replacement procedure, with overlap only where supported and appropriate |
This is a decision aid, not a substitute for your provider's incident procedures.
Verify two different outcomes
Retrieval test: After revoking Gazebo access, the old token or connection should no longer obtain a credential.
Provider test: After invalidating a provider key, verify that the provider rejects it using a safe, non-destructive check permitted by the incident process.
The first result does not imply the second. Gazebo retrieval records help establish key delivery; provider and runtime records help establish subsequent use.
Make offboarding routine
Named agent identities make it easier to remove retrieval access without editing every other profile. Review inactive profiles with their owners, but do not assume that no recent retrieval means no activity: a runtime may reuse a retained key.
Use separate provider credentials when the provider supports them and independent invalidation is important. Shared provider keys remain shared dependencies even with separate retrieval identities.
For the full response process, see AI agent incident response. For ongoing ownership and reviews, use the credential-management lifecycle.