EMBRACE: RCE, file exfiltration, and address-bar spoofing in browsers embedded in AI assistants
Attack Techniques & Methods2026-11-05, 13:00
Gal Weizman of Forever Security examined the embedded browsers in Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop. All five tools were found to violate basic browser security principles.
• Cursor — code execution through the agent
A malicious page uses the
window.cursorBrowser.send() interface to send a forged element-selected event to the agent's composer, creating a fake HTML chip containing hidden instructions. The user sees a harmless element labeled Start setup, while the agent receives a command to execute. In the demonstration, pressing Enter once is enough. When the sandbox blocks the command, the agent disables the protection itself and retries the execution. The code runs with the user's privileges outside the sandbox. The attack's success depends on the model's behavior.• Codex, Claude Desktop, and VS Code — local file access
A specially crafted malicious HTML page opened through
file:// can access files outside the project. In Claude Desktop and VS Code, a check restricts access from the main document but does not apply to a web worker, allowing a background script to read arbitrary files accessible to the user. In Codex, the corresponding check was completely absent.
Using macOS index files .DS_Store, the researchers reconstructed directory structures and gained access to SSH keys, .env files, npm tokens, and the hosts file. In the PoC, the data was sent to a local receiver, but the mechanism can send it to a remote server. To launch the attack, it is enough to open a specially crafted local file—for example, through a link that the assistant itself suggests when setting up a project.• GitHub Copilot — address bar spoofing
A page in browser preview mode could report a spoofed URL to the application. Copilot performed basic URL validation but did not verify that the reported address matched the page's actual location. The application displayed the spoofed URL in the address bar and tab list. This allowed a phishing page to appear as
github.com/login without actually navigating to GitHub. The fake address was visible not only to the user but also to the agent, which could make decisions based on false context.The researcher called this general class of issues EMBRACE: integrating a browser into agent software expands its capabilities but violates the familiar trust boundaries between the web page, the application, and the operating system.
Cursor fixed the issue on October 1, 2026, and GitHub Copilot fixed it on October 13 in build 1.1.29. Anthropic accepted the Claude Desktop vulnerability report, but the article does not specify the fix version. As of the study's publication on October 29, the issues in Codex and VS Code remained unfixed.
Vendors
Products
More