CVE-2026-66298: Livebook: sandboxed JS bypasses origin check, forces exec

Origin Validation Error vulnerability in livebook-dev livebook allows untrusted notebook output JavaScript to trigger session-wide keyboard shortcuts, including forced evaluation of all cells and runtime restart. Livebook's JS-view feature renders notebook-defined JavaScript inside a sandboxed, cross-origin iframe specifically because that JavaScript is untrusted. The trusted iframe shell in iframe/priv/static/iframe/v5.html forwards every keydown event fired in its own window to the parent page without consulting Event.isTrusted, so an event synthesized by the untrusted script through window.dispatchEvent is forwarded exactly as a genuine keystroke would be. The parent-side relay in assets/js/hooks/js_view.js reconstructs and re-dispatches it on the live page with no further validation, and because assets/js/hooks/session.js registers the global shortcut handler on the document in the capture phase, that handler acts on the replicated event regardless of how it was produced. Sandboxed output JavaScript can therefore drive Livebook's session-wide keyboard shortcuts. Two of them reach LivebookWeb.SessionLive and execute immediately with no confirmation: the shortcut for queueing full evaluation runs every cell in the notebook, and the shortcut for reconnecting the runtime disconnects and reconnects it, discarding in-memory state. A third shortcut deletes the focused cell behind a confirmation dialog that the user can permanently dismiss, after which it too executes silently. Forced full evaluation is the significant consequence, because it causes the notebook's own Elixir code to run without the user choosing to evaluate anything. A user who merely opens a notebook obtained from a third party, or reached from published documentation, can have its code executed on their runtime. Livebook also mirrors cell outputs to every connected client, so a malicious output triggers in a collaborator's browser as soon as it renders. This issue affects livebook: from 0.5.0 before 0.18.7 and from 0.19.0 before 0.19.9.

An adversary crafts a Livebook notebook (.livemd) containing a JS-view cell whose output JavaScript, once rendered, calls window.dispatchEvent to synthesize a keydown event matching Livebook's 'evaluate all cells' shortcut. They publish or share this notebook via a blog post, GitHub repo, Slack channel, or internal wiki page framed as a useful ML/data-science example. A victim opens the notebook in their own Livebook instance (or is a collaborator connected to a session where someone else opens it, since outputs are mirrored to all clients); the moment the JS-view output renders, the sandboxed iframe forwards the forged keydown to the trusted parent page, which re-dispatches it on the document in capture phase, triggering the global shortcut handler with no confirmation prompt. Every cell in the notebook now executes automatically, running any attacker-authored Elixir code as the victim's Livebook runtime — potentially exfiltrating data, pivoting to connected systems, or corrupting a legitimate ML pipeline running in that session — and the attacker can additionally force a runtime reconnect to wipe in-memory model/data state as a secondary disruption.