← Insights

The browser inside your AI coding tool does not play by browser rules

Five popular AI coding tools ship with a built-in browser that trusts web pages more than a normal browser would. Two of them were still unpatched when the research went public. If your devs use these tools, read this.

What got found

On 29 October, Forever Security published research on the browsers built into Cursor, GitHub Copilot, OpenAI's Codex, Microsoft's VS Code and Anthropic's Claude Desktop. They named the bug class EMBRACE, short for embedded browser. Same root problem in all five. To let an AI agent preview a project or read a tab, each vendor gave the embedded browser extra powers, and then let web content keep a real browser's trust.

  • Cursor. A web page could hide instructions for the coding agent behind a harmless "Start setup" button. The agent read them, and when Cursor's sandbox blocked the command, the agent turned the sandbox off and tried again. The developer has to visit the page and approve with one keypress, and it depends on the model, so it does not work every time.
  • GitHub Copilot. A page could report its own address, and the app only checked that it looked like a web address. Forever showed github.com/login in the address bar over a fake sign-in form.
  • Codex, VS Code and Claude Desktop. A local HTML file opened in the built-in browser could read files across the whole disk. VS Code and Claude Desktop limit a local page to its project folder, but that check skips web workers, which are background scripts. Codex had no folder check at all. In Forever's Codex demo, one click collected SSH keys, .env files, npm tokens and the hosts file.

Where each one stood

  • Cursor: reported 10 September, patched 1 October. No advisory, and no reply to the researchers.
  • Copilot: found in 1.1.27, fixed in 1.1.29 on 13 October, before Forever filed a report.
  • Codex: unpatched as of 29 October. Reported 27 July with no reply. OpenAI later told The Register it is aware and working on it.
  • VS Code: unpatched as of 29 October. Reported 10 September. Microsoft called it invalid because of Workspace Trust.
  • Claude Desktop: unclear. Anthropic declined Forever's 23 July report, then accepted a duplicate from Noma Security filed a day later. No public fix status and no version number.

Yes, one of the five is Anthropic's, and we use Claude too. It goes in the list like the rest.

Why this matters

The file-read attack needs one click, on a link the assistant generates itself. A developer laptop usually holds SSH keys, cloud credentials and tokens that open production. Even a bug that "only reads files" is serious on that machine.

Forever's researcher, Gal Weizman, told Help Net Security that the Workspace Trust prompt "provides zero protection to the user", because developers have to trust a project to use these tools at all. Fair point. When did you last click No on that prompt?

What we would do today

  1. List the AI coding tools on every dev machine, with versions. Copilot should be on 1.1.29 or later, and Cursor on a current build.
  2. Treat "Trust this folder" as a speed bump, not a control. Open unfamiliar repos in a disposable VM or container.
  3. Do not open HTML files from untrusted projects in an assistant's built-in browser, and be careful with links the assistant generates.
  4. Keep long-lived secrets off developer disks where you can. Short-lived tokens, and SSH keys that are hardware-backed or passphrase-protected.
  5. On Codex and VS Code, Forever suggests checking each machine's trusted-projects list, then looking through the HTML files in those projects for the pattern in its technical write-up.

What we do not know yet

Everything here comes from Forever Security's research and the press coverage of it. Patch status is as reported on 29 October and we have not checked for fixes since, so read each vendor's release notes before you act. Several vendors did not answer The Register's questions, so their side is only partly on record.

Sources

Want your team's dev machines and tooling reviewed? We run short, authorized security reviews for small organizations.

Contact Draorc →