IPCRIME Breach Response Center | November 21, 2026. You may have given an AI assistant access to an airline account, email inbox, calendar, shopping account or online files. That does not automatically mean the account has been compromised. But if the assistant was unauthorized by the provider, if you shared a password with a service you no longer trust, or if it made an action you did not approve, you should review and revoke that access. This checklist distinguishes the type of connection before deciding what to change.
First: determine how the assistant connected
- Delegated access: if you approved the connection through the account provider’s normal authorization screen, look for that provider’s connected-apps or authorized-apps settings.
- Password sharing: if you typed your username and password into an unrelated agent or service, assume the provider may not be able to revoke access by disabling one application grant.
- Browser or device control: if the agent used your active browser session, inspect both the account session and any browser extension or local application controlling it.
- Unknown connection: stop giving the assistant new sensitive tasks and use the real account provider’s security controls to investigate.
Revoke the right kind of access
For a delegated connection, remove the app’s grant from the provider’s authorized-applications page and check whether the account still shows active sessions. For a shared password, change the affected password from a trusted device, review other services using the same password and invalidate sessions where the provider permits. For a browser-controlled agent, disconnect its browser access, examine installed extensions and sign out of sessions that may have been accessed unexpectedly. Not every platform supports every control; follow its specific instructions rather than assuming one password change undoes everything.
Check what happened while it had access
- Review bookings, cancellations, purchases, loyalty point use, forwarding rules, calendar changes and document sharing as relevant to the particular account.
- Inspect any new MFA devices, recovery phone numbers, alternate emails and linked applications; secure the main recovery email before other accounts.
- Compare important actions with receipts, emails or records directly in the provider’s own website rather than trusting only the assistant’s explanation.
- Keep a private timeline of unexpected actions and save transaction confirmations. Avoid uploading passwords or sensitive records to a public forum for help.
If you suspect a malicious agent or impersonation
An unexpected account-recovery call, an unfamiliar authorization screen, or a request to share a verification code may signal social engineering. Do not give codes or remote-access permissions to someone claiming to be support without verifying them through the service’s known channels. A request to move money to a ‘safe’ account is particularly suspicious. Use the account or financial institution’s official support channel for any disputed action.
A safer way to reconnect in the future
Prefer provider-supported authorization that grants narrow, time-limited permissions rather than direct password sharing. Give read-only access to research tasks, and require explicit confirmation for purchases, transfers, deletion, or changes to identity and recovery settings. OWASP’s agentic security resources highlight tool misuse and access-control risks; those concerns become much easier to manage when permissions are bounded and actions are observable.
What not to assume
An airline restricting unauthorized AI bookings is not evidence your agent stole credentials. A login from a new IP address is not by itself proof of fraud, and an account with no recent suspicious actions is not a guarantee that a broad token is harmless forever. Revoke unneeded access proportionately, monitor relevant activity and keep recovery under your control.
Sources
OWASP — Agentic AI security risks; NIST — account authentication guidance; United Airlines — November 11 policy announcement; IPCRIME Digital Arsenal — safer permissions guide. This page is independent practical guidance, not an allegation that a named AI product has been compromised.