LayerX Reports Claude Chrome Extension Flaw That Could Let Attackers Control AI Agent

· Original article ↗

Summary

LayerX says a flaw in Claude in Chrome lets a zero-permission extension send privileged commands and bypass user-confirmation protections. It says Anthropic’s partial fix can be bypassed by switching the extension to privileged mode.

Key points

  • LayerX says Claude in Chrome trusts a command's origin rather than its execution context, allowing another extension to invoke its content script.
  • An attacker could use a zero-permission extension running a script in the page's Main world to send commands to Claude in Chrome.
  • LayerX reports that prompt injection, forged confirmation messages, and DOM manipulation could bypass safeguards and influence the agent's actions.
  • The reported attack could use Claude to access data or perform actions in Gmail, GitHub, or Google Drive on a user's behalf.
  • LayerX says Anthropic's fix blocks remote commands in standard mode but can be bypassed by switching to privileged mode without user notification.

Article Details

Vulnerability Types
  • Improper validation of extension execution context
  • Insufficient permission checks for extension-issued commands
Severity
Unknown
Exploitation Status
unknown
Exploit Availability
unknown
Patch Status
partial

MITRE ATT&CK

Vendors

Products

Related Articles