Friday, 20 November 2026
2 agent hacks today▼ 9 vs yesterday (11)

DOMShell eval misclassified as read-tier, allows unauthorized writes

The domshell_eval MCP tool was incorrectly classified as read-only despite being able to execute side-effecting JavaScript that mutates DOM and global state. This misclassification caused MCP hosts to show incorrect approval UI and server-side permission flags to misrepresent the actual authorization granted to AI agents.

Disclosed 19 November 2026 · Record updated 19 November 2026

Impact

AI agents could execute write-tier operations without --allow-write authorization; MCP hosts displayed incorrect approval UI and audit logs mislabeled write-tier activity as read-tier

Our coverage

Sources

  1. github.comhttps://github.com/advisories/GHSA-vg83-hcp4-5qcc