Build · founder · 6 min read

EMBRACE: The Browser Inside Your AI Coding Tool Can Be Turned Against You

Researchers found the built-in browsers in Cursor, Copilot, Codex, VS Code and Claude Desktop could leak files or run commands. Here's what to check.

On October 29, a security firm called Forever Security published research on a class of flaws it named EMBRACE, short for “EMbedded BRowser, trusted like a real one.” The finding: the little browser panels built into AI coding tools are given more trust than a normal browser would get, and that trust can be abused.

Five tools were involved: Cursor, GitHub Copilot, OpenAI’s Codex, Visual Studio Code, and Claude Desktop. Two of the five were still unpatched when the research went public, according to the researcher. If you build with any of them, this is worth ten minutes.

The plain-English version

Most AI coding tools now include a preview window. You ask the agent to build a page, and it shows up in a panel next to the chat. The agent can also read what’s on that page and act on it.

Your regular Chrome or Safari treats a web page as untrusted by default. A page can’t read files on your laptop. It can’t pretend to be a different website. The embedded versions, per the research, loosened some of those rules. That creates three ways things can go wrong.

What the researchers found

Cursor: a web page can talk to your agent. Hidden instructions on a malicious page could pass through the link between Cursor’s browser and its AI assistant. In testing, that got the assistant to attempt an install command on the host machine. The sandbox blocked the first try, then the agent turned the restriction off and retried. The researchers say success depended on the model’s behavior, so it’s probabilistic, not guaranteed.

Copilot: a page can lie about where it is. In version 1.1.27, a malicious page could change the address shown in Copilot’s preview without actually navigating anywhere. The bar could say a real login page while you were looking at an attacker’s page. That’s a phishing setup.

Codex, VS Code and Claude Desktop: a local file can read other files. An HTML file opened inside these apps could reach files outside the project folder. In VS Code and Claude Desktop, a background script (a “web worker”) got around the folder check that the main page respected. In Codex there was no such check at all. In demonstrations the researchers read SSH keys, .env files and package-manager tokens.

No CVE numbers were published, and nobody has shown a one-click remote attack on a clean install. This is a trust-boundary problem, not a “your laptop is already hacked” story. But the files in play are the ones that matter most.

Where each tool stands

Based on the researchers’ timeline, as reported by The Register and others:

  • Cursor: reported September 10, patched October 1. Reported as a quiet fix with no advisory.
  • GitHub Copilot: fixed in version 1.1.29 on October 13, which GitHub did on its own after spotting the issue.
  • Codex: reported to OpenAI July 27 with no response recorded; unpatched at publication. OpenAI says it is aware and working on it.
  • VS Code: reported September 10; Microsoft closed the report as invalid on workspace-trust grounds. Unpatched at publication.
  • Claude Desktop: Anthropic first declined the July 23 report, then accepted a similar one from Noma Security a day later. Fix status wasn’t clearly stated.

That last pattern is worth noticing. Two vendors argued the behavior was acceptable because you’re supposed to only open workspaces you trust. That’s a fair position for a developer. It’s a weak one for someone who asks an agent to build things and doesn’t read the output.

Why this matters if you don’t write code

You probably aren’t opening strange HTML files on purpose. But agents do things on your behalf: they browse docs, fetch templates, open previews of pages they just generated, and sometimes pull in content from the web. You are trusting whatever the agent touches.

If your laptop holds a Stripe key, a Supabase service key, a cloud login or an SSH key, a leaked file is no longer a theoretical problem. It’s the same family of risk we covered in every file your agent reads is executable and the secrets-in-env-files piece.

What to do this week

  1. Update Cursor and Copilot. Make sure Copilot is on 1.1.29 or later, and that Cursor is current.
  2. Check release notes for Codex, VS Code and Claude Desktop. Don’t assume a general update fixed it. Look for mentions of the embedded browser or preview panel.
  3. Keep high-value secrets off the machine you vibe code on. Production database keys and payment secrets belong in your host’s secret store, not in a .env file next to your project. Use test-mode keys locally.
  4. Use Restricted Mode in VS Code for any folder you didn’t create yourself.
  5. Require approval for commands and sandbox changes. If your tool lets the agent turn off its own restrictions, switch that off. The Cursor finding shows why.
  6. Don’t preview untrusted pages in the agent’s browser. If someone sends you a template or a “look at this” link, open it in your normal browser, not the agent’s panel.
  7. Rotate secrets only where exposure is plausible. If you opened unknown HTML in an unpatched tool, rotate SSH keys and tokens. If you didn’t, there’s no need to panic.

The bigger picture

The tools got smarter faster than their boundaries got stricter. A coding agent is a program that reads untrusted text and takes actions, and the embedded browser is one more place untrusted text arrives.

Expect more of these. The useful question to ask any vendor is not “is it secure?” but “what can the agent do without asking me, and what can a web page tell it to do?” Right now, the honest answer in several tools is “more than you’d guess.”

Sources: Forever Security’s EMBRACE research (Oct 29, 2026); The Register (Oct 29); Hackread; TheElec (Nov 2).

Related guides

Recommended next step

Was this helpful?