Quick answer: When an AI agent or its credential may be compromised, block further retrieval and stop or isolate the runtime where possible. If a provider key may have been exposed, restrict, revoke or rotate it at the provider too. Use retrieval, runtime and provider records together to assess impact; retrieval revocation alone does not establish containment.
This is a technical response playbook, not legal advice. Assign an incident owner and involve the relevant service, security and legal contacts as needed. Do not wait for a complete forensic picture before taking necessary containment actions.
1. Identify what may be compromised
Separate the agent's retrieval token or OAuth connection from the provider credentials it can obtain. An exposed retrieval token may have been used to fetch more keys; a compromised runtime may already hold them.
Capture the detection time, affected identity, relevant service references and known symptoms without copying secret values into tickets or chat. Preserve available evidence while containing the incident.
Useful starting signals include provider alerts, unexpected API activity, secret-scanning findings, unusual retrieval attempts and reports from the runtime owner. None alone establishes the complete incident window.
2. Contain retrieval, runtime and provider access
Block new delivery. Revoke the affected Gazebo access or connection. Verify that the old authentication can no longer retrieve credentials.
Stop ongoing execution. Where you control the runtime, stop or isolate it and suspend relevant jobs. Consider whether it can restart, delegate work or use another credential source. Preserve forensic evidence through your incident process.
Invalidate exposed provider credentials. If a key or the runtime holding it may be compromised, restrict, revoke or replace that key at its provider. Identify legitimate consumers of shared keys and coordinate replacement without unnecessarily delaying containment.
Gazebo returns the actual stored provider key to the authorised runtime and its host. Revoking a Gazebo profile does not erase that value or prevent a retained copy from calling the provider.
Check other persistence. Depending on permissions, an attacker may have created another account, token, deployment or scheduled job. Invalidating the original key does not necessarily remove those paths. Provider and runtime investigation is needed.
For the distinction between these actions, see API key rotation versus revocation.
3. Build an evidence-based timeline
Keep observations separate from assumptions:
| Evidence source | What it can help establish | Important limitation |
|---|---|---|
| Gazebo credential-access records | Which authenticated agent requested a key, when and with what outcome | Not the provider calls made afterward |
| Provider activity records | Recorded operations, resources and provider identity | Shared keys may make agent attribution difficult; coverage varies |
| Runtime and job records | Tool execution, process activity and task context | May be incomplete, sensitive or affected by the compromise |
| Configuration and access history | Intended permissions and known changes | Current configuration may differ from the incident-time state |
Ask which credentials were released, what their provider permissions were, what actual activity is recorded, and where evidence is missing. A missing log entry does not prove no access occurred. Retrieval frequency is not API-call volume.
Read-only permissions can still expose sensitive data or proprietary information. Do not classify an incident as minor solely because writes were blocked. Consider the data, recipients, duration and possible onward disclosure.
4. Assess notification duties in parallel
Escalate promptly to the appropriate response and legal owners. Do not wait until the investigation is complete to determine whether a notification clock has started.
Under GDPR Article 33, a controller must notify the competent supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of a personal data breach, unless it is unlikely to result in a risk to individuals' rights and freedoms. A processor must notify the controller without undue delay. Scope, awareness, risk and the organisation's role require assessment; a leaked key alone does not settle every question.
For HIPAA-covered situations, HHS breach-notification guidance distinguishes recipients and reporting rules. Individual notification is generally required without unreasonable delay and no later than 60 days after discovery of a reportable breach. Reporting to HHS has different rules for breaches affecting fewer than 500 individuals, and media obligations depend on the circumstances. Do not treat “60 days” as permission to wait or as one deadline for every recipient.
Other laws, contracts and incident-reporting obligations may apply. Retrieval logs alone cannot establish what personal or health data was accessed, whether an exception applies or whether notification is required. Record decisions, uncertainties and the advice received.
5. Restore access with verified boundaries
Before restarting, address the compromise path and review provider permissions, retrieval policies and runtime trust. Replace exposed bootstrap credentials and provider keys as required. Avoid restoring the same broad access simply because the agent's task is urgent.
Where high-risk operations need human approval, implement it in the runtime or request-executing system that actually controls those operations. Gazebo's approval of credential retrieval is not approval of a subsequent refund, deployment or data export.
Test permitted work, denied retrieval and relevant provider restrictions using safe test cases. Confirm old credentials no longer grant the access you intended to remove.
6. Close with owners and follow-up evidence
Document what was known, what was contained, which provider credentials were invalidated, what remains uncertain and who owns the remaining work. Review related agents for the same exposure path rather than assuming the issue is isolated.
Gazebo contributes named retrieval identities, release policies and credential-access history. It does not supply complete provider forensics or guarantee one-action containment. For ongoing reviews and offboarding, use the credential-management lifecycle.