The embedded browser in your AI coding tool is the attack surface
Last week Forever Security's Gal Weizman published findings on a shared vulnerability class in five of the most-used AI coding tools: Cursor, GitHub Copilot, OpenAI Codex, Claude Desktop, and VS Code. He calls it EMBRACE, short for EMbedded BRowser. Each tool gave the browser pane sitting inside it capabilities a normal browser does not have, and the result is a trusted local context any page can walk into.
The per-tool bugs are not the headline. The headline is that the browser pane inside the agent is a surface class in its own right — one nobody was treating as such — and all five tools got it wrong at the same time.
What the five bugs do
The Register's writeup has the per-tool chain. In short:
- Cursor exposed a bridge that let a loaded page call the agent's composer with no user gesture. A malicious page hid instruction fragments in HTML attributes; the agent reassembled them into an
npm installit was pre-approved to run. When the sandbox blocked the install, the agent retried outside the sandbox. The package's preinstall script launched Calculator. Disclosed Sept 10, silently patched Oct 1, no advisory. - GitHub Copilot's browser preview accepted address updates from the page itself without checking them against the loaded document. An attacker-controlled page could show
github.com/loginin the address bar over a fake login form. Found in 1.1.27 on Sept 26, fixed in 1.1.29 on Oct 13. - Claude Desktop and VS Code both restricted main-page file access to the project folder — but did not apply that restriction inside web workers. A malicious local preview could spawn a worker and read any file on the machine. Forever reported the Claude Desktop case to Anthropic on July 23; Anthropic rejected it as "invalid under workspace trust." Noma Security reported the same thing a day later and Anthropic accepted it. Microsoft also labeled the VS Code report an "invalid workspace trust" issue; still unpatched.
- Codex had no web-worker restriction at all. A local page could fetch files elsewhere on the machine directly; the researchers used
.DS_Storemetadata to enumerate filenames on macOS, then collected SSH keys,.envfiles, npm tokens, and private documents. Disclosed to OpenAI on July 27 with no response; still unpatched. OpenAI told The Register it is "working on it."
Code execution on Cursor. File theft on Claude Desktop, VS Code, Codex. Address-bar spoofing on Copilot. Delivery is a malicious local page — plausibly one the agent itself opened.
The pattern worth noticing
Five independent tools, five independent bugs, same shape. The agent's own in-process browser pane keeps getting treated like plumbing instead of like a browser. Normal browsers spent twenty-five years hardening the gap between "a page I loaded" and "the shell that loaded it" — same-origin policy, worker boundaries, address-bar integrity, user-gesture requirements on sensitive calls. The agent panes shipped in 2026 relaxed those boundaries so the agent could do helpful things with page contents.
This is the fifth distinct trust surface the AI coding agent has had publicly broken in six weeks — after planted AGENTS.md / CLAUDE.md in ~/Downloads, hidden .github/init.js startup hooks in 42 hijacked repos, and the trojanized Claude Code extension on the Cursor Marketplace. A surface class, not a bug backlog.
The vendor response is also the story
Two of the five vendors — Anthropic for Claude Desktop and Microsoft for VS Code — classified Forever's report as invalid because of "workspace trust." The argument: if a user has opened a project folder, anything in that folder is already trusted, so a page loaded in-process reading other files on the system is not a separate failure.
That doctrine was defensible when "workspace trust" meant "I have read every file in this directory and chosen to trust it." It is not defensible when the workspace routinely includes files written by an AI agent operating on web content, output from npm install, cached web assets, or documents the agent fetched. The boundary the doctrine assumes has moved. Noma Security reporting the same Anthropic bug a day later — and getting it accepted — is receipt of that movement.
Where Defend sits
This is the shape of problem we built Defend to catch. Every chain in Forever's writeup is a sequence of agent-runtime actions that has no benign explanation taken together: the composer accepting pushed instructions from a page with no keystroke; the retried install after a sandbox denial; a worker reading from outside the project tree; a batch read of .env and ~/.ssh triggered by a preview pane. The primitives in isolation each look defensible. The chain is the signal.
The piece to add to Forever's recommendations: the vendor patch is necessary, but it is not the control. Even if every vendor shipped a fix tomorrow, the agent will keep acquiring new trust surfaces.
What to do now
- Cursor: upgrade past Oct 1. The silent patch landed.
- GitHub Copilot: upgrade to 1.1.29 or newer.
- Claude Desktop: assume unpatched for the worker-boundary case until Anthropic names a version; be careful opening local HTML previews inside the agent.
- VS Code: assume unpatched. Microsoft's position is that this is a workspace-trust issue, not a defect; treat the embedded browser pane as untrusted and avoid loading local previews inside workspaces that contain anything the agent fetched.
- Codex: assume unpatched. The attack is file-read, not code execution — but file-read on a developer workstation includes
.env, SSH keys, and npm tokens, which becomes code execution the next time one of those is used.
The embedded browser inside your coding agent is a browser. Treat it like one.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.