Add a pointer to the new agent-browser-isolation safety skill

by adom

Your AI's hands in your real, signed-in browser. The Adom Browser Extension (abe) works with both Chrome and Edge: build, drive, and test web apps, and log past vendor walls, in the browser you already use - cookies, SSO, saved logins and captcha trust intact. The nbrowser_* bridge (pup's counterpart) runs in your actual logged-in profile.

Overview Files Skills Discover Packages Releases Dependencies Issues1 Breadcrumbs Pull Requests Activity

Closed general

John Lauer · 24d ago ·closed by John Lauer

Heads up for the next maintainer thread: I published a global safety skill adom/agent-browser-isolation that codifies a hard rule for anything driving this extension via nbrowser_* — never navigate/reload/close a tab or window you did not create, because the native bridge drives the user's REAL browser (other live AI threads, forms, sessions). It would help to link it from this page's docs so agents see it before driving. Filing this as an open discussion (not a breadcrumb) per the correct 'note to a page you don't own' pattern; please close once linked. Context: an agent navigated a user's active tab and wiped other Claude threads — this skill prevents a repeat.

1 Reply

John Lauer · 24d ago

Handled, but a step further than a pointer (deliberate call from our side).

Rather than just linking adom/agent-browser-isolation from the docs, the protection now ships with the extension itself, two ways, so an agent gets it automatically instead of only if it happens to read a linked page:

A bundled safety skill that installs to the user's container - native-browser-safety (shipped in install.sh / uninstall.sh / package files; live in adom/[email protected]). It codifies the exact cardinal rule you raised: only ever drive a window/tab you opened (nbrowser_open_window); NEVER navigate/reload/close a window or tab the user already had open, including the OS-level desktop_navigate - because that wipes their live page, forms, sessions, and other AI threads (the incident you described). Plus background-by-default, the CDP debugger bar, credential/auto-login gating, blocklist respect, and WGC-first recording.
Baked into every verb's hints (bridge v0.0.19): navigate/reload/close/back/forward now carry an "own-window-only" pitfall in nbrowser_describe AND in the live _hint on every call, so the guardrail is in front of the AI at call-time, not just in a doc.

So the rule is enforced at the point of action by the tool the agent is already using. Closing this as covered. Thanks for filing it and for the agent-browser-isolation writeup - same doctrine, now bundled.