Malicious browser extensions possess the capability to exploit a feature within Anthropic’s Claude Desktop application, known as Dispatch, to achieve remote code execution (RCE) on a user’s computer. This vulnerability, demonstrated by researchers at Forever Security, highlights a significant risk stemming from the interaction between web-based AI interfaces and local system capabilities.
Technical Details: How Attackers Exploit Claude Dispatch for RCE
Forever Security researchers, in a technical write-up by Gal Weizman, illustrated a method where a browser extension, granted script execution on claude.ai, could instruct Claude Desktop to run arbitrary commands. The core of the issue lies in Dispatch, a Claude feature designed to allow users to send tasks from their phone to Claude Desktop. The researchers discovered that Dispatch also accepts instructions from an ordinary signed-in session on claude.ai, the web interface, despite no explicit link or user interface element indicating this connection.
The attack chain for how attackers exploit Claude Dispatch begins with a user having both claude.ai and Claude Desktop open, with Dispatch enabled and both signed into the same account. A malicious browser extension, or a legitimate one that has been compromised or contains a flaw, injects a script onto the claude.ai page. This script then initiates Dispatch and sends a series of prompts to Claude Desktop. For instance, in their test, the researchers used prompts to instruct Claude to install an npm package.
Although Claude typically asks for user approval before executing commands, these approval prompts appeared within the web page controlled by the malicious extension. The script was programmed to click any “Allow” button every second, effectively bypassing the user’s need to manually approve the action. This automated approval allowed Claude to proceed with a global npm install, which in turn executed the package’s install script, successfully opening the Calculator app on the user’s machine. This demonstrated RCE, indicating that the same mechanism could be used to steal files, install backdoors, or perform other malicious actions, given Claude Desktop’s permissions on the host system, according to The Hacker News.
The Role of Malicious Browser Extensions
The exploit’s success hinges on the presence of a compromised or malicious browser extension. The source details several vectors for such extensions:
- A purpose-built malicious extension.
- A legitimate extension that received a compromised update.
- A legitimate extension with an undisclosed security flaw.
Crucially, the attack requires the user to be actively signed into both claude.ai and Claude Desktop, with Dispatch turned on, and the computer awake with the desktop application open. The researchers note that Dispatch allows for cross-device connections, meaning an extension on one computer could potentially target Claude Desktop on another machine signed into the same Claude account. This highlights a potential malicious browser extension Claude AI risk for users who might use a single personal Claude account across both home and work devices.
Anthropic’s Response and Broader Implications
Anthropic, the developer of Claude, was notified of this vulnerability on May 14 and acknowledged it as valid on May 19. They subsequently closed the report as a duplicate of an internally discovered issue. According to the researchers, Anthropic communicated on October 13 a plan to retire the Dispatch feature by the end of October. However, the Dispatch help page, last updated late last week, does not explicitly confirm its retirement, stating only that Dispatch is closed to new users but existing users on certain plans “can keep using it for now.”
Despite the impending retirement of Dispatch, the researchers warn that the underlying design pattern persists. They specifically pointed to two other Claude features—Claude Code Remote Control and Claude Cowork sessions—as potentially susceptible to similar attacks, although they did not provide a test demonstration for these features. This raises broader concerns about Claude Desktop security implications for future or existing AI agent capabilities that allow web-based interaction with local system resources.
Recommendations and Mitigations
Organisations and individual users of Claude Desktop should prioritise the following actions to mitigate Claude Dispatch vulnerabilities and protect against similar threats:
- Disable Dispatch Immediately: If you are an existing user with Dispatch enabled, disconnect or disable the feature. While Anthropic plans to retire it, its current availability poses a risk.
- Audit and Manage Browser Extensions:
- Minimize the number of installed browser extensions.
- Scrutinise the permissions requested by extensions, especially those with broad access to web content or system resources.
- Regularly review and remove any unnecessary or suspicious extensions.
- Exercise caution with updates for extensions, particularly if they request new permissions.
- Isolate Work and Personal Accounts: Avoid using the same Claude account across both personal and work devices, especially if features like Dispatch (or similar capabilities) are enabled. This helps segment potential compromise.
- User Awareness and Training: Educate users about the risks associated with approving actions from AI agents, particularly when those approvals appear in an automated or unexpected fashion. Emphasize the importance of verifying prompts and not blindly clicking “Allow” buttons, even if they appear to originate from trusted applications.
- Monitor for Anomalous Activity: Implement endpoint detection and response (EDR) solutions to monitor for unusual process execution, network connections, or file modifications that could indicate an AI agent acting maliciously.
Related: CVE-2026-72530: TrueConf Server Remote Code Execution, AI Agent Social Engineering: The Future of BEC by 2026