This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
For most of the history of online reputation, the thing to manage was a page. A news article, a review, a search result, a Wikipedia entry. You could find it, read it, and respond to it. Anthropic’s announcement on Tuesday points to a different kind of surface: a session on somebody else’s computer.
What Anthropic released
Alongside an upgraded Claude 3.5 Sonnet, Anthropic introduced computer use in public beta. Through the API, developers can direct Claude to operate a computer “by looking at a screen, moving a cursor, clicking buttons, and typing text.” Asana, Canva, DoorDash, Replit and The Browser Company are among the companies Anthropic says are already exploring it.
Anthropic is unusually frank about the limits. It calls the capability “experimental,” “at times cumbersome and error-prone,” and notes that scrolling, dragging and zooming still give the model trouble. On the OSWorld benchmark, Claude scored 14.9% in the screenshot-only category, ahead of the next-best system at 7.8%. Leading the field at under 15% says a lot about how early this is. The company also warns that computer use “may provide a new vector for more familiar threats such as spam, misinformation, or fraud,” and says it has built classifiers to detect harmful use.
So this is not a finished product. It is an early, honest signal of direction.
From reading about you to acting near you
Here is how I would describe the shift in the information chain. Media writes about a company. Search indexes and ranks it. Wikipedia condenses it. AI systems summarize all of the above. Each of those layers produces text a communications team can see.
An agent that operates a desktop adds a layer that produces behavior instead. It might fill out your contact form, compare your prices with a competitor’s, try your checkout flow, or gather information from your site for a report. It does this inside a session on a machine you do not own, guided by instructions you never see, using screenshots of your pages rather than a reader’s patience.
That changes several familiar assumptions.
Your interface becomes evidence. If an agent cannot complete a task on your site because a pop-up hides the button, the report it hands back may simply say your company “did not provide” the information. A design choice turns into a factual claim about you.
Attribution gets murky. When a person has a bad experience, there is usually a complaint, a review or a social post. When an agent fails, the record sits in a developer’s logs. You may never learn that the failure happened, only that a customer formed a view.
Impersonation gets cheaper. Anthropic’s own list of risks, spam and fraud, describes exactly the kind of activity that can run under a brand’s name: fake sign-ups, scripted posts, automated outreach that looks like it came from you or your customers.
What to do while it is still early
None of this calls for alarm. A 14.9% benchmark score suggests the volume will stay small for a while. But the direction is clear enough to prepare.
Make sure core tasks on your site can be completed through a plain, visible interface, without hidden steps that only work for a patient human. Ask your security and digital teams whether they could tell the difference between unusual automated sessions and normal traffic, and who would be told. And when you assess how your company is represented, add a question to the usual ones about search results and AI answers: if a capable assistant tried to do something on our behalf, or about us, what would it find, and what would it report?
Reputation has long been about what people read. It is starting to include what software does when nobody is watching the screen.