Let’s talk about agentic browsing for a minute, because a lot of otherwise technically literate people are missing the actual problem.
Openclaw is an easy example, but it’s not the only one. Perplexity’s products raise many of the same concerns, and this pattern shows up all over the space.
The issue is not just “bad security policy” so much as it is the underlying design.
These systems work by exposing browser secrets directly to the agent context: passwords, cookies, session state, local storage, and other authenticated data. That may be the simplest implementation path, but it is also the most reckless.
And the moment you do that, you bypass the browser’s normal security and privacy boundaries. You are no longer asking a browser to act on your behalf inside its own security model. You are extracting protected state and handing it to an LLM pipeline.
And that means you may be injecting your most sensitive data into third-party inference providers that have no meaningful legal obligation to keep it private from internal review, retention, monitoring, or future policy changes.
This is the last user behavior we should be interested in normalizing.
If you wouldn’t hand your passwords and session tokens to a shady guy in a trench coat on the street corner because he offered you candy, don’t hand them to an inference pipeline either.
Basic common sense, y’all.
Be safe out there.