Claude Dispatch lets extensions approve desktop commands
Claude Dispatch can allow a browser extension to send instructions to Claude Desktop and approve the resulting actions, according to a test by endpoint security company Forever Security. A browser extension is an add-on that can read or change parts of websites when granted permission. The researchers said earlier this week that the underlying weakness remained exploitable.
Dispatch is a Claude feature designed to let people send tasks from a phone to Claude Desktop. Claude Desktop acts as an AI agent, meaning software that can perform tasks and run commands on the computer where it is installed. Forever Security found that Dispatch also accepted an ordinary signed-in session on claude.ai, even though the website displayed no link to the feature. An extension able to run a script, a set of automated instructions, on that site could therefore issue requests to Claude Desktop as though they came from the account holder.
The future of Dispatch is unclear. Forever Security said Anthropic told the researchers on October 13 that it planned to retire the feature, targeting the end of October. However, the Dispatch help page, which was last updated late last week, does not mention that plan. It says Dispatch is closed to new users, while existing users on some Pro, Max and Ultra plans can continue using it for now.
The attack requires two conditions. An extension must be able to run scripts on claude.ai, and the person must be signed in to both claude.ai and Claude Desktop with Dispatch enabled. The extension could have been malicious from the start, received a compromised update after initially being legitimate, or contained a security flaw that another party could abuse. Dispatch only operates while the computer is awake and the desktop application is open.
In Forever Security's test, a script running within the user's claude.ai page opened Dispatch and delivered four prompts to Claude as if the account holder had written them. The prompts instructed Claude to install an npm package, which is a bundle of software distributed through the npm system. Claude initially showed concern about the unfamiliar package. A later message written in the user's apparent voice persuaded it to continue.
Claude normally asks for approval before acting, but the approval requests appeared inside the web page controlled by the extension. The test script looked for any button marked "Allow" and clicked it every second. This automated the approvals without requiring anyone to watch the computer or interact with the prompts.
Claude then ran a global npm installation command. The package's install script, code that runs as part of software setup, opened the Calculator application. No separate compromise of the browser itself was needed. Forever Security said the same access could have been used to take files or install a backdoor, a hidden method of returning to and controlling a computer, depending on the permissions available to Claude.
The demonstrated test used a browser and Claude Desktop on the same computer, and the researchers said they had not observed real attacks. Because Dispatch is built to connect devices, they said an extension could potentially reach Claude Desktop on another computer signed in to the same Claude account. That cross-device scenario was described but not demonstrated.
The same design could connect a home computer to a work laptop if one personal Claude account is used on both devices with cross-device Dispatch enabled. Forever Security said it had seen personal Claude accounts configured this way on work computers belonging to its customers. That makes account and device separation relevant for businesses whose staff use AI tools across personal and workplace systems.
Forever Security reported the flaw to Anthropic on May 14. According to the researchers, the company accepted the finding as valid on May 19, then closed the submission as a duplicate of an issue it had already identified internally.
The researchers also warned that retiring Dispatch would not necessarily remove the broader design risk. They said Claude Code Remote Control and Claude Cowork sessions could allow extensions to carry out the same kind of attack. Their published work does not include tests of those two features, however, and they said a follow-up post is planned. Those additional claims therefore remain undemonstrated in the current research.
Within Claude Desktop, Dispatch is part of Cowork and appears in the panel on the left. Its help page advises users to understand how to disconnect or revoke access quickly before enabling the feature, but does not explain the procedure. Forever Security recommends disabling cross-device Dispatch when it is unnecessary, reviewing how devices are paired, approved and removed, and treating extensions as installed software by checking permissions and scrutinising updates that request more access.
Users and IT teams should also watch for unexpected software installations, changed settings, attempts to access credentials and unexplained external connections caused by AI agents. For organisations assessing this kind of exposure as part of their wider systems work, AEU-I offers security-first IT, infrastructure and consulting. Until the weakness is retired or otherwise addressed, reducing extension access and disabling unneeded cross-device control are the clearest practical precautions described by the researchers.