Unpatched flaws in Codex and VS Code expose developers to attack
Researchers found vulnerabilities in five AI developer tools that use embedded browsers, which can give web content powers that ordinary browsers restrict. The flaws could enable attackers to run commands on developers’ computers, read files or mislead users about the site they are viewing; researchers say vulnerabilities in OpenAI’s Codex and Microsoft’s VS Code remain unpatched.
The affected tools are Cursor, GitHub Copilot, Claude Desktop, VS Code and Codex. Cursor patched a code-execution flaw on 1 October after disclosure on 10 September, while GitHub fixed Copilot’s address-bar spoofing issue in a 13 October build without receiving the researchers’ report. The researchers said local pages could reach files in Claude Desktop and VS Code through browser workers, while Codex allowed direct file access; Anthropic disputed the Claude report on workspace-trust grounds.
In short
- Embedded browsers gave web content unsafe capabilities.
- Codex and VS Code flaws were reported as unpatched.
- GitHub and Cursor issues were fixed.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
The researchers’ findings raise a legitimate security concern: embedded browsers can give malicious web content access to files or command execution beyond what users expect from an ordinary browser. Until the reported Codex and VS Code flaws are fixed or independently shown not to be exploitable, developers may reasonably want clear disclosure, timely patches and safeguards against untrusted pages.
The case against
The reports describe potential attack paths, but their practical risk depends on how the tools are configured and what trust users have granted; Anthropic, for example, disputes its finding on workspace-trust grounds. Security claims should be assessed tool by tool and confirmed with vendors, since the cases differ and at least two issues were already fixed, including one without a report from the researchers.
AI Art Culture Research Science Software Technology
Read the full article at the source →
Originally published by The Register as “Researchers find bugs in 5 AI developer tools, with OpenAI and Microsoft flaws still unpatched”.