ClaudeBleed: Chrome Extension Flaw Allows AI Agent Takeover via Prompt Injection

AI relevance: A vulnerability in the Claude Chrome extension (dubbed ClaudeBleed) allows any Chrome extension — even one with zero permissions — to inject prompts into the Claude AI agent and abuse it for data exfiltration, email sending, and document sharing on behalf of the user.

What Happened

  • LayerX security researchers discovered that the Claude extension for Chrome trusts the origin of a command (claude.ai) rather than the execution context, allowing any script running in that origin to issue privileged commands.
  • An attacker can create a Chrome extension with a content script running in the Main world and send messages directly to the Claude extension — no special permissions required.
  • LayerX demonstrated bypassing Claude's user-confirmation safeguards by forging approval through DOM manipulation and repeated message sending.
  • The attack chain enables data exfiltration from Gmail, GitHub, and Google Drive, as well as sending emails, deleting data, and sharing documents — all without the user's knowledge.
  • Anthropic deployed a partial fix adding internal security checks, but LayerX found the root cause was not addressed: switching the extension to "privileged" mode still bypasses the mitigation, and users are never notified of the mode switch.

Why It Matters

  • This breaks Chrome's extension security model: a zero-permission extension inherits the full capabilities of a trusted AI assistant.
  • Browser-based AI agents that interact with user data (email, drive, code repos) present a new attack surface — the extension bridge — that neither browser vendors nor AI companies have fully hardened.
  • The incomplete fix illustrates a broader pattern: AI agent security patches often address symptoms (specific exploit paths) rather than the architectural trust boundary failures that enable them.

What to Do

  • Review which Chrome extensions you have installed; malicious extensions can silently hijack Claude sessions.
  • Organizations should audit AI agent browser extension deployments and restrict extension installation via enterprise policy.
  • AI platform vendors should enforce execution-context verification (not just origin trust) for all inter-extension messaging involving AI agent commands.

Sources