Anthropic unveiled the Claude plugin to control the Chrome browser + video

Anthropic has unveiled a new Chrome browser plugin that allows the company’s artificial intelligence assistant to take control of the browser and perform complex tasks automatically. This moves a more serious entry into the competitive and high -risk “Agentic AI”; An area where artificial intelligence systems can interact directly with computer user interfaces.

The new “Claude for Chrome” plugin allows users to order artificial intelligence to do things in the browser; For example, you can tell artificial intelligence to plan or look at your emails and look for specific items. This system can see what is displayed on the screen, click the buttons, fill the forms, and move the websites.

Anthropic, however, has announced that it has only provided the plugin for only 3 trusted users and in a limited beta version, citing serious security concerns, especially the Prompt Injection attacks.

Claude Artificial Intelligence Plugin for Chrome

The new Chrome plugin can transform organizational automation. Instead of sophisticated APIs, these artificial intelligence agents can work with any software that has a graphical user interface.

However, in its internal experiments, the anthropic has found a very worrying security vulnerability. In this type of attack, called “Prampt Injection”, a destructive command can embed hidden commands into a website, email or document to deceive artificial intelligence and force the harmful actions without the user’s knowledge.

Self -on -anthropic tests showed that these attacks in the 4.9 % They have been successful. For example, anthropic researchers found that a malicious email, which had set himself as a security guidelines, ordered Claude to remove all user emails “for mailbox health”, and artificial intelligence did so without any confirmation.

Anthropic has implemented several protective layers to deal with these risks; Including website level access permits, mandatory confirmation for high -risk measures such as purchasing or sharing personal data and blocking access to sensitive sites such as financial services. These measures have been able to reduce the success rate of Prapete injection attacks from 4.9 percent to 4.9 percent, although the company itself acknowledges that this is not enough for public supply.

The anthropic cautious approach to presenting this plugin is in full contradiction with the more aggressive movements of competitors. OpenAI has previously released its agent called Operator for ChatGpt Pro users; Microsoft has also merged similar features into its Copilot Studio platform.