EMBRACE: Two of Five AI Coding Tools Still Unpatched

On October 29, 2026, Forever Security researcher Gal Weizman published the EMBRACE research, showing that the preview panes embedded in five AI coding tools — Cursor, GitHub Copilot, OpenAI Codex, VS Code and Claude Desktop — were being trusted as if they were standalone browsers, violating a core security rule that ordinary browsers enforce. The class, named EMBRACE for an embedded browser trusted like a real one, stemmed from vendors widening permissions in their built-in browsers, giving them an elevated level of trust. The research covered five issues in total, and two tools remained unpatched at publication.

According to the findings, a visited web page could execute commands on the local development machine through the AI editor, while a local HTML file could read SSH private keys. Both issues coexisted across all five tools until October 29, 2026. The report frames the problem as an attack surface between the browser component and the local dev machine. The researchers shared steps for auditing one's own machine and applying temporary mitigations, and advised users to check vendor release notes, confirm that a fix has landed, and re-verify afterwards.