Why it matters
The Hacker News reports that a rogue extension able to script claude.ai can forge a click that triggers allowlisted Claude for Chrome tasks. In unattended mode, the task can run without approval and access Gmail, Google Docs, or Calendar data.
My takeaway: The failure is in the extension’s trust and permission boundary, not the language model. Browser agents should reject synthetic gestures, initialize in approval-required mode, constrain cross-extension access, and never infer user intent from page-controlled events.