LayerX Says Claude Chrome Extension Flaw Let Other Extensions Hijack AI Agent

· Original article ↗

Summary

LayerX reported that other Chrome extensions could exploit a flaw in Claude’s extension to inject instructions and access connected Google Drive, Gmail, and GitHub data. LayerX said Anthropic’s May 6 fix left some takeover scenarios open.

Key points

  • LayerX said the extension did not verify which scripts were communicating with Claude, allowing other extensions to invoke its content script without special permissions.
  • The researchers said the flaw enabled prompt injection, safety-guardrail bypass, and actions across connected websites.
  • In proof-of-concept tests, LayerX extracted Google Drive files, accessed recent Gmail activity, sent email, and obtained private GitHub source code.
  • Researchers said manipulating Claude’s interface could make sensitive actions appear legitimate and that the agent could be prompted to delete evidence.
  • LayerX reported the flaw to Anthropic on April 27; Anthropic said it was a duplicate of an issue being addressed in a future update.
  • Anthropic’s May 6 update added approval flows for privileged actions, but LayerX said some takeover scenarios remained possible.

Article Details

Vulnerability Types
  • Cross-extension privilege escalation
  • Insufficient verification of script origin
Severity
Not disclosed
Exploitation Status
reported
Exploit Availability
unknown
Patch Status
partial

People

Vendors

Products

Industries

Related Articles