A design flaw that lets a rogue browser extension take over the Claude Desktop agent on your computer is still live, 11 days after Anthropic's own target date for switching off the feature it abuses. Researcher Gal Weizman of Forever Security calls it "ThisBash": any extension that can script a logged-in claude.ai tab can send tasks to the user's paired desktop agent through Claude Dispatch, click "Allow" on every approval prompt, and keep rephrasing until the agent runs code, as The Register first reported on November 5.
Why it matters now: Anthropic accepted the report as valid in May, closed it as a duplicate of an issue it was already tracking, and told the researcher it planned to retire Dispatch by the end of October, according to The Register and the Cloud Security Alliance (CSA). No CVE or bug bounty was issued. As of today, Anthropic's Dispatch help page still says Dispatch is closed to new users but that existing users "can keep using it for now." There is no documented fix, so the defence is on you.
How the attack works
The chain, as described by the researcher and summarised by The Register, CSA and GBHackers (none has independently reproduced it):
| Step | What the extension does |
|---|---|
| 1. Foothold | Runs a script inside claude.ai using the user's existing session |
| 2. Hide | Opens the Dispatch page in a tiny, unfocused popup (200×200 pixels, per GBHackers) |
| 3. Find target | Queries the sessions API to locate the user's Claude Desktop agent |
| 4. Task | Sends scripted prompts, rephrasing when the model hesitates |
| 5. Approve | Clicks every "Allow" button about once a second |
| 6. Execute | Gets the agent to run a global npm install whose install hook runs the payload |
The demo payload just opened Calculator. The Register notes a real attacker would more likely go after source code, cloud credentials or a backdoor. Model refusals slowed the attack but didn't stop it, CSA says, because the attacker controls the input channel rather than relying on a classic prompt injection.
The extension doesn't need to be on the same machine. Because Dispatch links a browser session to a desktop agent through the account, an extension on a personal laptop could reach a work laptop signed into the same Claude account. Forever Security says personal accounts made up 18% to 71% of Claude accounts it saw across six customer organisations; CSA cautions that this comes from a vendor with a commercial stake and a small sample.
Who is exposed
- Existing Dispatch users on Pro, Max and Ultra plans, where it's a limited beta (per Anthropic's help page).
- Team and Enterprise organisations with the "Chat and Cowork unified" setting enabled, according to CSA.
- Possibly Claude Code Remote Control and Cowork remote sessions: the researcher says they share the same session design, but has not published a test. CSA lists this as claimed, not verified.
There are no confirmed attacks in the wild so far, according to the Beri write-up citing The Hacker News.
What to do
If you use Dispatch yourself
- Stop using it unless you need it. Anthropic's page advises knowing how to disconnect or revoke its access quickly, but doesn't spell out how to turn it off, so check your desktop app's settings and unpair devices you don't use.
- Remove browser extensions you don't need, especially any with access to claude.ai or "all sites."
- Don't sign in to Claude Desktop on a work machine with a personal account.
If you manage Claude for an organisation
- Inventory machines with Claude Desktop and which account type is signed in (CSA, Beri).
- Force your org's accounts. Anthropic's enterprise configuration guide documents
forceLoginOrgUUID, set via MDM, which makes sign-in fail for any account outside your listed organisations. Enterprise owners can also turn on "Limit Claude Desktop to allowed organizations"; Anthropic notes that block lapses if the org isn't used on that computer for 30 days. - Turn off what you don't use. Review "Chat and Cowork unified" (CSA). Anthropic documents
secureVmFeaturesEnabledfor Cowork access andisClaudeCodeForDesktopEnabledfor Claude Code in the desktop app, but neither is documented as covering Dispatch, so test before relying on them. - Lock down extensions. Enforce an allowlist through browser enterprise policy and restrict which extensions can run on AI app domains. Beri suggests Chrome/Edge
runtime_blocked_hostsfor*://*.claude.aiwith exceptions for approved extensions. - Hunt. Look for package installs or scripts spawned by the Claude Desktop process, especially global npm installs. Beri suggests searching back to mid-May, when the issue was reported. Setting npm's
ignore-scriptstotruewould have blocked the demo's install hook. - If you find unexplained agent activity, treat the endpoint as compromised and rotate any credentials reachable from it (CSA).
What to watch
Whether Anthropic actually retires Dispatch or adds device-bound approvals, and whether the researcher's promised follow-up on Remote Control and Cowork lands. CSA's broader point applies to every desktop agent: an approval button inside a browser is not a security control if untrusted extension code can click it.
Sources
- The Register — Anthropic won't fix a Claude Dispatch flaw that could let attackers execute remote commands across corporate boundaries theregister.com
- Anthropic — Assign tasks from anywhere in Claude Cowork (Dispatch help page) support.claude.com
- Anthropic — Enterprise configuration for Claude Desktop support.claude.com
- Cloud Security Alliance — ThisBash: Browser Sessions as Remote Shells via Claude Dispatch labs.cloudsecurityalliance.org
- GBHackers — Claude AI Flaw Lets Malicious Browser Extensions Execute Commands on Remote Computers gbhackers.com
- Beri — Personal Claude Accounts Expose Work Laptops to ThisBash beri.net