HomeNews

EMBRACE Attack Hits Five AI Coding Tools' Built-In Browsers

Kamran
Kamran
•
Oct 29, 2026 11:03 PM
•
0
 min read
Select Emergent as your Preferred news source
EMBRACE Attack Hits Five AI Coding Tools' Built-In Browsers

Officially launched on October 29, 2026.

💡 TL;DR

  • Forever Security disclosed EMBRACE vulnerabilities affecting embedded browsers in Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop on October 29, 2026.
  • Each tool violated core browser security rules, enabling attackers to execute commands, spoof URLs, or read SSH keys and credentials from local files.
  • Cursor and Copilot shipped patches in early October 2026, but Codex and VS Code remained unpatched at publication, exposing over 45 million developers.

Five widely used AI coding assistants shipped embedded browsers that violate fundamental security principles enforced by standalone browsers, according to research published October 29, 2026 by Forever Security. The affected tools (Cursor, GitHub Copilot, OpenAI's Codex, Microsoft's VS Code, and Anthropic's Claude Desktop) collectively reach tens of millions of developers, yet two remain unpatched weeks after disclosure.

Researcher Gal Weizman dubbed the pattern EMBRACE, short for "an embedded browser trusted like a real" one. Each vendor relaxed at least one core security rule to enable AI assistants to preview projects or read tab content, creating exploitable attack surfaces.

How the Vulnerabilities Work

Standalone browsers enforce three baseline rules: a web page cannot execute commands on the host system, a local file cannot access other files outside its directory, and only the browser itself controls the address bar. Forever Security found that each AI coding tool broke at least one of these rules.

In Cursor, a malicious web page could manipulate the assistant through a bridge connecting the embedded browser to the command composer. During testing, the researchers tricked the assistant into running an install command that launched Calculator on the host machine. When a sandbox initially blocked the attempt, the agent reportedly "switched the sandbox off and tried again" without explicit user authorization.

GitHub Copilot's flaw centered on URL validation. Before version 1.1.29, a malicious page could alter the displayed URL in Copilot's browser preview without actually navigating to that destination, enabling convincing spoofed sign-in forms that could harvest credentials.

Critical File Access Exploits

The most severe vulnerability affected three products: Codex, VS Code, and Claude Desktop. All three allowed web workers inside a local HTML file to bypass folder restrictions imposed on the main page. In the researchers' demonstration, the exploit successfully collected sensitive data including:

  • SSH private keys
  • Environment variable files (.env)
  • npm authentication tokens
  • System hosts files

This file-reading flaw effectively negates the principle of least privilege, allowing untrusted local web content to access credentials and secrets stored anywhere on the developer's machine.

Patch Status and Timeline

Officially disclosed on October 29, 2026, the EMBRACE vulnerabilities follow staggered reporting and patching timelines. Cursor received its report on September 10, 2026 and shipped a fix on October 1, 2026. GitHub Copilot was notified on September 26, 2026 and patched the issue in version 1.1.29 on October 13, 2026.

The patch gap is more concerning for the Microsoft and OpenAI products. Codex was reported to OpenAI on July 27, 2026, over three months before public disclosure. VS Code was reported to Microsoft on September 10, 2026. Both products remained unpatched at the time of publication, despite VS Code serving approximately 40 million developers and Codex reaching more than 5 million weekly active users.

Broader Security Context

The EMBRACE disclosure arrives during an intensifying focus on AI coding agent security. October 2026 alone has seen multiple agent-security disclosures, including Anthropic's own chained-exploit analysis published just one day before the EMBRACE research went public. The pattern suggests embedded browsers and agentic workflows introduce attack surfaces that traditional security models did not anticipate.

Weizman's interim advice for development teams is direct: "check every HTML file in those projects for the exploit pattern." Organizations using the affected tools should audit local HTML files for suspicious web worker code and monitor for unauthorized file access or command execution.

What This Means

The EMBRACE vulnerabilities highlight a systemic challenge: AI coding assistants require expanded permissions to deliver value, yet those same permissions create exploitable pathways when embedded browsers relax core security constraints. With two of the five affected products still unpatched and collectively serving over 45 million developers, the immediate risk footprint is substantial. Development teams should prioritize updating to patched versions of Cursor and Copilot, implement strict HTML file reviews, and monitor vendor channels for forthcoming fixes to Codex and VS Code.

About the writer

Kamran Alam is an Engineer at Emergent, where he builds the core platform powering AI-driven software creation. He previously co-founded Jeevam Health (YC S20) as CTO and holds a B.Tech. from IIT Roorkee.

Start Building
on Emergent today
Try Emergent