Claude Dispatch: unpatched flaw lets browser extensions run code
A browser extension can use Claude Dispatch to make Claude Desktop run code, even on another PC. Anthropic has not fixed the flaw, reported in May.

A browser extension that can act on the claude.ai page is able to use Claude Dispatch to make Claude Desktop run commands: on the user’s computer or on another machine connected to the same Claude account. No browser exploit is needed and user interaction is minimal. The technique, called ThisBash, was discovered by Gal Weizman of Forever Security and published as an exclusive by The Register on 5 November 2026.
All the technical details come from The Register’s article, which had access to the researcher’s report and video demonstration: DataBreaches.net republished an excerpt and The Hacker News picked up the story on 6 November. Anthropic did not answer the publication’s questions.
What happened
Dispatch is the Claude feature that lets users send and manage tasks for Claude Desktop, the agent that runs commands on the computer, from their phone or browser. According to Weizman, the Dispatch interface and API were reachable from an ordinary authenticated web session on claude.ai: an extension with access to that page could therefore send instructions to Claude Desktop as if it were the legitimate user.
According to The Register, Anthropic acknowledged the report as a valid vulnerability in May, but closed it as a duplicate of a known issue; it paid no bounty, assigned no CVE, and told the researcher it would retire Dispatch at the end of October. That did not happen: as of 30 October, the Claude support page listed Dispatch in limited beta on some Pro, Max and Ultra plans, not available to new users but still usable by those who had already enabled it.
Weizman argues that the problem stems from a design choice, not just from Dispatch: in his view, the same sessions shared between web and desktop also exist in Claude Cowork (now available from the web version) and in Claude Code Remote Control. The researcher says he told Anthropic, which reportedly decided not to act.
How the attack works
Prerequisites: an extension with access to claude.ai, either malicious or legitimate but hijacked (Weizman mentions supply chain compromise and the black market for extensions), and a user signed in both to claude.ai and to Claude Desktop with Dispatch enabled. With Cowork and Remote Control, according to the researcher, enabling Dispatch is becoming less and less necessary.
The chain described in the report has six steps:
- the extension registers a content script on claude.ai in the MAIN world, sharing the page’s JavaScript and the user’s session cookies;
- the script opens Dispatch (
https://claude.ai/cowork/agent) in a small unfocused popup, the only visible trace of the attack; - using the session cookies it calls
GET /v1/sessions?tags=cowork-dispatch-localand obtains the session ID of the agent running on the computer; - it writes instructions into the message box (force prompting) and tells the agent to act;
- an automatic click every second approves any confirmation request from Claude, so the chain does not stop if nobody is in front of the screen;
- if the agent hesitates, the script rephrases and insists until it gets code execution.
In the proof of concept the agent runs npm install -g on a package with a preinstall.js script that opens the calculator. In a real attack, The Register notes, the package could steal source code and cloud credentials or install a backdoor.
Indicative MITRE ATT&CK mapping, derived from the reported facts: T1176.001 (Software Extensions: Browser Extensions), T1195.002 (Compromise Software Supply Chain), T1550.004 (Use Alternate Authentication Material: Web Session Cookie), T1059 (Command and Scripting Interpreter).
How far it reaches
If the controlled desktop is the same machine as the browser, the attack is an escape from the browser sandbox. But Dispatch works across computers: the compromised browser can become the entry point to an agent on another machine. The most sensitive case is a personal Claude account used both on a home PC and on a company laptop: an extension in the home browser ends up running commands on the work computer.
Forever Security told The Register that it often sees accounts shared between home and office: in six customer organisations, the share of personal Claude accounts in corporate environments ranged from 18% (255 out of 1,233) to 71% (20 out of 28).
Impact for European and Italian companies
The risk affects those who have introduced Claude agents on workstations, often without an inventory, and those who tolerate personal accounts on company devices. For NIS2 entities, Article 21 requires measures on supply chain security, cyber hygiene and access control: browser extensions and AI agents with permissions on the operating system fall within this scope. If the agent can read files containing personal data, health data included, abuse is also a potential breach under the GDPR.
What to do now
- Take an inventory of workstations with Claude Desktop and check where Dispatch, Cowork and Claude Code Remote Control are in use; turn Dispatch off where it is not needed.
- Separate accounts: no personal Claude accounts on company devices and no company accounts on personal devices.
- Govern extensions: allow only approved ones and, with Chrome and Edge enterprise policies (
ExtensionSettingswithruntime_blocked_hosts), prevent extensions from acting on claude.ai. - Limit the possible damage: users without administrator rights, no long-lived credentials on workstations running agents, and consider
ignore-scripts=truein the npm configuration to block install scripts. - Monitor processes started by Claude Desktop: global npm installs, shells and outbound connections to unusual destinations.
- Follow Anthropic’s communications on the retirement of Dispatch, bearing in mind that, according to the researcher, the problem remains in Cowork and Remote Control.
How ARXDOME helps
ARX Sentry keeps an inventory of the software and processes on every computer, so you know where Claude Desktop is installed and in which version, and collects security logs and events from your machines: suspicious executions launched by the agent reach the SOC pages, and file integrity monitoring flags changes in the folders you have chosen.
Last updated:


