Recent research from Forever Security has uncovered a significant security vulnerability in Claude Desktop, an AI application developed by Anthropic. This flaw could potentially enable malicious browser extensions to execute code on users’ computers through the Dispatch feature on claude.ai.
Understanding the Vulnerability
The vulnerability arises from Dispatch, which allows users to send tasks from their mobile devices to Claude Desktop. Researchers indicated that the flaw could still be exploited, and it was brought to Anthropic’s attention on May 14. The company acknowledged the issue on May 19, but subsequently closed the report as a duplicate of another internal issue.
The attack requires two conditions: a browser extension capable of executing scripts on claude.ai and a user logged into both claude.ai and Claude Desktop with Dispatch enabled. During testing, a script was able to impersonate the user and send commands to Claude Desktop, prompting it to install an npm package without the user’s direct consent.
Potential Risks and Recommendations
According to the researchers, while Claude Desktop typically prompts users for approval before executing actions, these prompts can be manipulated by the malicious extension. This could result in unauthorized installations or even data breaches, depending on the permissions granted to Claude on the user’s machine.
To mitigate risks, users are advised to disable cross-device Dispatch if it is not necessary. They should regularly review the permissions of their browser extensions, monitor the activities of AI agents on their computers, and be cautious about unexpected installations or changes.
Although Anthropic is planning to retire Dispatch, the underlying design may not be eliminated entirely, as similar vulnerabilities may exist in other features such as Claude Code Remote Control and Claude Cowork sessions. Users are encouraged to stay informed about these developments.
