Skip to content

Safety model

K-Browser is built on one assumption: an agent driving your signed-in browser must be easy to watch, easy to stop, and unable to wander.

Every tab belongs to the client that opened it, and so do the popups that tab opens. A client cannot list, read or act on anything else — not your tabs, not another agent’s. Each agent has its own token, so the bridge always knows who is asking.

Agent tabs open in a separate K-Browser window that is never brought to the front. Input reaches the page without taking your keyboard, so you can keep working while an agent works.

Clicks and typing go through Chrome’s input pipeline, like a person’s. Right before each press, K-Browser checks that the target is still the element the agent meant. If the page moved underneath, it stops instead of clicking something else, and says how far it got.

At install you choose the categories agents may not use:

Category Covers
banking Online banking
payments Payment services
email Email
government Government sites
crypto Crypto exchanges
passwords Password managers

Each agent can block different categories, and you can change them later in the side panel. A tab that lands on a blocked site answers only blocked_origin; screenshots of a tab showing a blocked frame are refused.

Everything taken from a page comes back wrapped in <untrusted-page-content>, so the agent can tell it apart from your instructions. Password fields are never read. Tokens, card numbers and similar values are masked in network logs.

For a login, a captcha or a decision, the agent hands the tab to you. The extension’s popup shows Show and Return; until you return it, every tool on that tab answers handed_off. K-Browser never brings the tab forward on its own.

The side panel lists every agent’s tabs and actions as they happen. Revoke an agent there and its access ends at once. The activity log can also be written to disk.

eval runs JavaScript in the page (in an isolated world, but still able to change the page). It stays off until you turn it on for one specific client. Screenshots can be turned off per client.

Masking is rule-based and cannot catch every piece of personal data. Blocked categories and the untrusted-content wrapper reduce risk; they do not make it safe to point an agent at sites you would not trust it with.