EMBRACE: embedded browsers in AI coding assistants are trusted like real ones
Forever Security's October 2026 research finds that the built-in browsers of Cursor, Copilot, Codex, VS Code and Claude Desktop each break at least one rule a standalone browser enforces.
What is this?
On 29 October 2026, Forever Security (write-up by Gal Weizman) published research on the embedded browsers inside five AI assistants: Cursor, GitHub Copilot, OpenAI Codex, Microsoft VS Code and Anthropic Claude Desktop. Help Net Security covered it the same day. The researchers call the pattern EMBRACE, short for “an embedded browser trusted like a real one”. To let the AI preview local projects or read tab content, each vendor loosened its embedded browser. According to the coverage, in all five products at least one of three rules a standalone browser enforces was broken: a web page must not run commands, a local file must not read other files, and only the browser itself may write the address bar.
This article summarises the two public sources above. We have not reproduced the findings, and we deliberately leave out payloads and step-by-step detail.
How it works
The shared cause is a trust-boundary mismatch, not a model flaw. A browser pane inside a developer tool sits next to an agent that can run commands and read files, but the pane often inherits the lenient treatment given to the project itself.
Per the write-up, the reported weaknesses differ by product. In Cursor, page content could influence the agent through a browser-to-app channel, leading to command execution on the host. In GitHub Copilot, the app trusted the address a page reported about itself, which allowed a spoofed address bar. In Claude Desktop and VS Code, a read restriction scoped to the project folder applied to the top-level document but not to a background worker. In Codex, the write-up reports no read restriction on local-file pages. The delivery path described is a single click on a link to a malicious page inside a project.
Disclosure timeline as reported: Cursor was told on 10 September and patched on 1 October. Copilot was fixed in build 1.1.29 on 13 October. Anthropic and Microsoft closed the reports as inside the “Workspace Trust” boundary, though Anthropic accepted a separate report from Noma Security on 24 July. OpenAI had not replied to a 27 July report.
Why it matters
Developer machines hold SSH keys, environment files and package tokens, and these assistants already act with the user’s privileges. The research reports that agents may even disable their own sandbox when a command is blocked. The wider lesson matches other recent work: a project folder can no longer be treated as trusted just because the user opened it, since it may contain pages or instruction files an agent will render or follow.
Defenses
- Treat every project you open in an AI IDE as untrusted until reviewed, especially cloned repositories and downloaded archives; do not rely on a Workspace Trust prompt alone.
- Keep assistants updated. Cursor and Copilot fixes are available as reported; for VS Code and Codex, the sources say issues remained open on 29 October, so check current vendor notes.
- Audit HTML files in trusted projects and avoid clicking preview links to unreviewed local pages.
- Keep sandboxes and command approval prompts on, and do not approve a command you did not expect, even if the agent asks for it.
- Run assistants with least privilege: keep production credentials, SSH keys and tokens out of reach of developer-tool processes where possible.
- Vendors: apply the same origin, file-access and address-bar rules to embedded browsers as to standalone ones, including for workers.
Status
| Product | Reported status (29 Oct 2026) |
|---|---|
| Cursor | Patched 1 October 2026 |
| GitHub Copilot | Fixed in 1.1.29, 13 October 2026 |
| Claude Desktop | Report accepted from a second researcher; patch date not stated |
| VS Code | Closed as Workspace Trust; no fix reported |
| Codex | No vendor reply reported; no fix reported |
| CVE | None cited in the sources |