Isn't there a situation where the agentic browser, acting correctly on behalf of the user, needs to send Bitcoin or buy plane tickets? Isn't that flexibility kind of the whole point of the system? If so, I don't see what you get by distinguishing between agentic and no agentic browsing.
Bad actors will now be working to scam users' LLMs rather than the users themselves. You can use more LLMs to monitor the LLMs and try and protect them, but it's turtles all the way down.
The difference: when someone loses their $$$, they're not a fool for falling for some Nigerian Prince wire scam themselves, they're just a fool for using your browser.
Or am I missing something?
You're right that if the user logs into a sensitive website, the "isolated browsing" mitigation stops helping. We don't want the user to accidentally end up in that state though. Separately, I can also imagine use-cases for agentic browsing where the user doesn't have to be logged into sensitive websites. Summarizing Hacker News front page, for one.