An AI agent can be talked into anything.
Anyone letting an AI agent work on sites holding personal data has this problem, from hospitals and banks to the companies that build agent browsers. Hidden text on a page can give an agent orders, an agent that was tricked may not say so, and no single site can defend against every agent.
The browser sees every page the agent opens and every form it sends, so it's the one place to enforce the rules. It learns the identifying details on a protected page, then blocks them from leaving in any form, using a small model and format rules that run inside the browser. It keeps its own log, which the agent can't edit, and only real clicks can change the permission slip.
If the browser, not the agent, holds a permission slip, personal data can't leave even when the agent is fooled. The slip says which sites the agent may use, what personal data it may use and where, and for how long and under what conditions. The browser acts as the safeguard, and ordinary browsing isn't disrupted.