A preview pane that could read the whole disk
Forever Security says the browsers built into Cursor, GitHub Copilot, OpenAI's Codex, VS Code and Claude Desktop each broke a basic browser rule. Cursor and Copilot have fixed theirs; Codex and VS Code were still unpatched on Thursday, the researchers said.
Matthew Field, CC BY-SA 3.0, via Wikimedia CommonsThe web browsers built into five AI coding tools, Cursor, GitHub Copilot, OpenAI's Codex, Microsoft's VS Code and Anthropic's Claude Desktop, each broke a basic browser security rule, the research firm Forever Security said on Thursday. Two of the flaws, in Codex and VS Code, were unpatched when it published, Help Net Security reported.
The coding tools embed a browser so the AI can preview a project a developer is building or read what is in a tab. To make that work, Forever says, each vendor loosened the rules an ordinary browser enforces, and in every case one of three broke: a web page cannot run commands on the computer, a local file cannot read other files, and only the browser can write the address bar. Forever calls the pattern EMBRACE, for an embedded browser trusted like a real one, in a technical write-up by its researcher Gal Weizman.
The file-reading flaw is the most serious. Claude Desktop and VS Code confine a local page to its project folder, but the check covers only the page itself, the researchers say: a web worker, a background script any page can start, falls outside it and can read any file the user can. Codex had no folder check at all. Forever's demonstration read macOS's hidden index files, which list the folders a user has opened in Finder, to map the disk, and on its test machine collected SSH keys, .env files holding application secrets, npm publishing tokens and the hosts file.
The other two flaws needed more help from the user or the model. In Cursor, a web page showed a "Start setup" button carrying a hidden instruction to the agent; one press of Enter approved it, and when Cursor's sandbox blocked the command, the agent switched the sandbox off and tried again. The test command opened the Calculator app. Help Net Security noted that the attack depends on the model and does not always work. In GitHub Copilot, the app checked only that a page's reported address looked like a web address, so Forever's page displayed github.com/login above a fake sign-in form, and the agent's own list of tabs repeated the false address.
| Product | Flaw, as the researchers describe it | Reported | Vendor response, per Forever | Status on 29 October |
|---|---|---|---|---|
| Cursor | A web page can get the agent to run a command | 10 September | No reply | Patched 1 October, 21 days later |
| GitHub Copilot | A page can show a false address, such as github.com/login | Found 26 September; no report sent | Fixed independently | Fixed in version 1.1.29 on 13 October |
| OpenAI Codex | A local page can read files across the disk; no folder check | 27 July | No reply | Unpatched, 94 days on |
| Microsoft VS Code | A web worker escapes the project-folder check | 10 September | Declined, citing Workspace Trust | Unpatched, 49 days on |
| Anthropic Claude Desktop | A web worker escapes the project-folder check | 23 July (Noma Security: 24 July) | Declined Forever's report, citing Workspace Trust; accepted Noma's as valid | Not stated, 98 days on |
The vendors' responses differ more than the flaws do. Cursor patched its flaw on 1 October, three weeks after Forever's report, without replying, and GitHub fixed Copilot's in version 1.1.29 on 13 October before Forever had filed anything, the researchers say. OpenAI has not responded to a report filed on 27 July. Microsoft and Anthropic declined Forever's reports as outside their security boundary, pointing to Workspace Trust, the prompt that asks a user whether to trust a project folder; Anthropic did accept as valid a report on the same issue that Noma Security filed a day after Forever's, according to the write-up, which does not say whether Claude Desktop has since been fixed. The accounts are the researchers'; Komposite has not reproduced the attacks.
Forever's argument is that the trust prompt does no work. "Not trusting a project inside an AI harness makes it unusable, so everyone just hits 'Trust'," Mr Weizman told Help Net Security, adding that "that prompt provides zero protection to the user." By our count, the Codex report has gone 94 days without an answer and the original Claude Desktop report 98 days without a stated fix. For teams using Codex or VS Code, Mr Weizman advised pulling the list of trusted projects from each machine's configuration file and checking every HTML file in those projects for the pattern the write-up describes.
Why it matters: the files the demonstration collected are the ones attackers use to move from a developer's laptop into a company's systems and software supply chain. When the hosting platform Railway thought environment variables had been exposed last week, it told every customer to rotate its keys, Komposite reported. Coding assistants are also now finding flaws in each other's makers: researchers using Claude were paid a $6,500 bounty by OpenAI in September.