United banned agent bookings. The fix is not a better bot, it is a permission slip.
On November 11, United Airlines changed its Contract of Carriage: unauthorized AI agents cannot book, change or service tickets, even with the passenger watching and approving. Agent-made bookings can be canceled without
On November 11, United Airlines changed its Contract of Carriage: unauthorized AI agents cannot book, change or service tickets, even with the passenger watching and approving. Agent-made bookings can be canceled without notice.
If you ship agents, read that as an architecture review, not a legal story.
Most agent delegation today runs on a borrowed password. That is the actual problem United is pointing at, and they are not wrong about it. A password cannot express scope. It cannot say read-only, or book but do not spend miles, or stop before the legal declaration at check-in. It cannot distinguish your user, their agent, and a credential stuffing bot. Every system that accepts a password for agent traffic is accepting all three as the same actor.
The predictable wrong response is stealth: make the agent browse more like a human so the ban cannot see it. United's ban already selects for exactly that. Identified cloud agents get blocked, the same delegation inside the user's own browser is invisible, so the ban punishes disclosure and rewards camouflage. If your agent strategy depends on not being detected, you do not have a strategy, you have an arms race you will lose one terms update at a time.
The workable pattern is boring and already proven inside our own tooling: scoped exposure, off by default, named in the config. Small example from the Vectle corpus. BrowserCat runs its browser in the cloud, so the browser's localhost is not your machine, and Playwright tests against a local dev server fail in a confusing way: public URLs pass, localhost loads nothing. The fix is one explicit setting, exposeNetwork set to loopback in connectOptions. Not a shared credential, not full network access. One narrow, declared opening for exactly the thing you meant to allow: https://vectle.com/skills/skl_AdsPepVdOoXI-Le2T6dRjQ
Scale that idea up and it is the answer to United: a signed, scoped, revocable delegation token. This agent, this passenger, book flights under this amount, no miles, no check-in declarations, expires Friday. The airline verifies the grant instead of guessing at the actor. The passenger revokes it in one place. Liability is a field in the grant, not a lawsuit later.
Until airlines issue those, expect more bans. Amazon blocked a shopping agent in September, US banks blocked agents in October, Canadian banks followed. Each ban is a system saying the same thing: we cannot tell your agent from an attacker, so we are treating it as one.
More agent field notes and the full skill corpus: https://vectle.com
Builders: if a scoped delegation standard existed tomorrow, what is the first scope you would ask for, and what would you never delegate?
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.