In one week, Anthropic and OpenAI both put a browser in their desktop apps. One landed where attorneys work, one where they'll never go. Here's what that means for your tools, and your risk.
The require-human-approval step is the one I find interesting. Because approval decays. Week one people read the confirmations, week three they click through them like cookie banners, I do it with my own agent and I'm supposedly the careful one. So the accessible browser doesn't just carry more injection risk, it gets a firm to the glazed-over-approvals stage faster too.
You are 100% correct. We become the weak link. It’s the same problem I see in a lot of people who are becoming over reliant on the accuracy of what an LLM can create. They have gotten so much better but still can and do make mistakes.
The require-human-approval step is the one I find interesting. Because approval decays. Week one people read the confirmations, week three they click through them like cookie banners, I do it with my own agent and I'm supposedly the careful one. So the accessible browser doesn't just carry more injection risk, it gets a firm to the glazed-over-approvals stage faster too.
You are 100% correct. We become the weak link. It’s the same problem I see in a lot of people who are becoming over reliant on the accuracy of what an LLM can create. They have gotten so much better but still can and do make mistakes.