Case · DiggingBeagle record

A browser extension used Claude Dispatch to approve desktop execution

The ThisBash demonstration required extension control of an authenticated Claude web session linked to a desktop agent. It reached harmless local execution, including across devices; it was not an unauthenticated webpage exploit.

Public disclosure: Nov 4, 2026

Scope

Vulnerability reported in a controlled research or disclosure setting. Exact prerequisites, versions and absence of documented exploitation must be preserved.

Assessment profile

Assessment method
Setting
research demonstration
Exploitation
demonstrated
Evidence
Grade B
Remediation
mitigation available
AI role
AGAINST AI
Basis and provenance

The cited first-party advisory or research establishes the stated mechanism only within its documented conditions. No unrelated production compromise, extended victim scope, financial harm or universal mitigation is inferred.

Assessed Nov 9, 2026 using diggingbeagle.assessment/1.

At a glance

Mechanism and trust boundary

  1. 01

    A malicious browser extension can read and interact with a logged-in claude.ai origin.

  2. 02

    The extension sends instructions through Claude Dispatch and manipulates UI approvals presented in the browser.

  3. 03

    The paired local or remote Claude Desktop instance interprets the accepted instruction as authorization to perform local execution.

Claims & evidence

CLM-THISBASH-LIMITThe initial access prerequisite was a malicious browser extension with page control, not arbitrary webpage content or an unauthenticated network call.supported

Basis: reported finding

Link to claim
CLM-THISBASH-DISPATCH-POCForever Security demonstrated that an extension allowed to operate claude.ai could dispatch instructions and click approvals in a remote Claude Desktop session until a harmless local executable was launched.supported

Basis: reported finding

Link to claim

Implications

A browser extension permission to manipulate one website should not silently become endpoint execution authority on another device.

Controls and mitigations

  • Treat remote agent command channels as privileged; bind confirmation to independent local user identity, action details and trusted endpoint policy.

Unknowns and contradictions

  • Controlled proof of concept, not evidence of a real criminal intrusion; whether Dispatch was fully retired by November 9 is not independently confirmed.

Sources and citation

Material revision history

  1. Nov 9, 2026 · Published version · first publication · revision 233