Notice: This article was created with AI.
What’s It About?
AI assistants can by now operate a browser in much the same way as a human: they open pages, click, scroll and fill in forms. The F.A.Z. tried this out with Claude. A plain instruction to end the Netflix subscription was accordingly enough to run through the entire cancellation process. Shortly afterwards the confirmation arrived by email.
The way back worked as well, according to the report: with a further command the cancellation was reversed. The precondition was that the testers signed in once in the Claude browser.
Background & Context
Technically, the AI reads the content of a page, plans the necessary steps and carries them out through clicks, text entries and scrolling. It therefore does not stop at chatting but acts as an agent. The piece describes AI agents as systems that pursue goals, gather information and carry out work steps independently.
As an example the F.A.Z. names Claude for Chrome, a browser extension from Anthropic. It can read page contents, click buttons, fill in forms, read out data and scroll through pages, including behind a login. Tasks are described in ordinary language, and every step can be followed in the open tab.
The fields of use named as typical are calendar and email management, data entry, recurring forms, expense reports and bookings. Websites can also be tested by running through user flows. According to the piece, developers can work with Claude Code in the terminal and have the result checked in the browser, including console errors and page state.
In one demonstration the AI first draws up a plan, for instance for searching for smartphones on eBay. The excerpt available breaks off at this point, which is why the exact course of the approval cannot be reproduced here. The piece also warns of security risks, above all of hidden instructions on websites and in emails that could lead an agent astray.
What Does This Mean?
- Routine tasks involving forms, such as cancellations or bookings, can be handed over to a browser agent in plain language.
- Anyone giving the agent access to logged-in accounts should watch the steps in the tab and deliberately confirm critical actions.
- Hidden instructions in web pages and emails count as the central risk. Caution is called for especially with sensitive accounts and payment processes.
Sources
Prompts für KI im Browser: Claude, bitte kündige mein Netflix-Abo (FAZ)
Video auf YouTube zum Netflix-Test mit dem Browser-Agenten Claude (YouTube)
Claude, bitte kündige mein Netflix-Abo – Video (FAZ)
FAZ-Beitrag auf X zum Netflix-Test mit Claude (X)
This article was created with AI and is based on the listed sources as well as the language model’s training data.
Further Reading: GPTs, Skills, Plugins, Agents — Who Offers What
Hands-on projects to build yourself