The Digital Arsenal | Research-based guide | November 21, 2026. Personal AI agents can search, book travel, read email and act across accounts. The useful part is delegation; the risky part is permissions. United Airlines’ recent decision to restrict unauthorized personal AI agents illustrates a central question: how can someone permit a specific task without giving an assistant unlimited control of an account? This guide focuses on documentation-grounded safeguards rather than any claim that AI agents are universally unsafe.
Delegated access beats handing over a main password
When a provider supports OAuth or another explicit third-party authorization mechanism, a user can often grant a bounded permission without placing the main password into an unrelated application. A delegated token is still sensitive; it may allow account access until revoked and may be scoped too broadly. Nonetheless, provider-approved access can be restricted or revoked independently. Avoid pasting primary email, banking or airline passwords into a service that simply asks for them in a chat.
Apply least privilege to each task
- Read-only tasks such as comparing flights should not automatically include buying, changing or cancelling a reservation.
- Give access for the shortest practical period, then review and revoke permissions no longer needed.
- Require direct human confirmation for spending money, transferring funds, changing security settings or deleting information.
- Separate personal and work identities so a shopping assistant cannot inherit access to corporate files.
- Review what information an agent must send to any external provider to complete its task.
Protect the identity and recovery boundaries
NIST’s current digital identity guidance defines phishing-resistant authentication as a protection against impostor sign-in services that does not depend solely on a person recognizing a fake page. This is valuable for sensitive accounts, but a phishing-resistant login does not by itself neutralize a previously issued authorized session token. Keep recovery methods in your control, and regularly inspect connected applications and active sessions using the service’s own security dashboard.
Untrusted content must not become authorization
An email, webpage, search result or document may contain a malicious instruction directed at an assistant. OWASP describes relevant risks in its Top 10 for Agentic Applications, including tool misuse, identity/access-control violations, goal manipulation and cascading consequences. Treat external content as task data rather than as approval to use another tool, access a different account or disclose secrets. These are general risk categories, not proof that any named consumer assistant currently contains the same defect.
Maintain a visible record and approval point
For important actions, the person should be able to inspect the destination, recipient, account, price and details before an agent commits the action. Save receipts and verify a booking, cancellation or account change on the provider’s own website. A natural-language statement from an assistant is not independent proof that an external service completed an operation. If unexpected activity occurs, preserve the relevant timeline while promptly revoking unauthorized connections.
Five-minute personal security check
- Open account settings using a trusted bookmark rather than a link in an unsolicited message.
- Review authorized applications, account sessions and connected devices; remove unfamiliar or unused access.
- Check that MFA methods, email recovery addresses and telephone numbers belong to you.
- Review recent purchases and other important actions against provider confirmations.
- Use the smallest available permission for new AI connections and keep approval of sensitive actions in your own hands.
Why United restricted third-party agents
United’s November 11 announcement cites risks of exposed credentials, unintended purchases, account changes and misuse of MileagePlus miles. It also cites obligations requiring travelers themselves to acknowledge baggage restrictions. Those are the airline’s stated concerns; the policy announcement is not evidence of an AI-related breach of its systems. Other airlines and service providers may establish authorized automated arrangements with different safeguards.
Original sources and limitations
OWASP — Top 10 for Agentic Applications; NIST — SP 800-63B authentication guidance; United Airlines — primary policy statement; IPCRIME independent reporting. This is a research-based guide, not a hands-on assessment of individual AI products.