Weaknesses in five AI coding tools allow code execution, file access, and URL spoofing

En bref

• Five AI-assisted development products are affected: Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop.
• The reported impacts include arbitrary code running, unauthorized local file access, and misleading browser address bars.
• Vendors named in the findings are Cursor, GitHub, Microsoft, OpenAI, and Anthropic.
• No CVE identifiers have been published, so patch tracking cannot rely on CVE feeds alone.

Five AI-assisted development products are in scope: Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop. The vendors named in the findings are Cursor, GitHub, Microsoft, OpenAI, and Anthropic. The reported weaknesses fall into three impact classes: arbitrary code running on the host, unauthorized copying or exfiltration of local files, and browser address bars being made to show misleading URLs.

Because these tools sit close to source code, terminals, and local file systems, a flaw that permits code execution or file access can matter more than a typical browser bug. Developers often grant broad workspace permissions and run generated commands without review. Address-bar spoofing adds a separate phishing risk: a user may trust a displayed URL that does not match the actual destination.

No CVE identifiers have been assigned, so defenders cannot rely on CVE-based patch tracking alone. The severity hint is informational, meaning the immediate priority is inventory and verification rather than incident response. Still, organizations should not dismiss the issues: AI coding assistants and desktop clients are increasingly installed outside formal software approval processes.

Watch for vendor advisories and product updates for all five tools. Confirm which versions are deployed, whether extensions or plugins are allowed, and whether desktop clients can read sensitive directories. Until fixes are confirmed, reduce auto-execution and broad file access where possible, and remind staff that address-bar text can be spoofed.

À faire maintenant

1. Inventory all installations and versions of Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop across managed and unmanaged endpoints.
2. Apply vendor updates as soon as they are available and verify that desktop clients and IDE extensions are on supported releases.
3. Disable automatic execution of generated commands or code where the product allows it, and require user review before running AI-produced actions.
4. Restrict workspace and file-system permissions for AI coding tools so they cannot read sensitive directories unless explicitly required.
5. Review allowed extensions and plugins, and block unapproved AI assistant integrations through endpoint or IDE policy.
6. Monitor outbound file transfers and unusual process activity from developer endpoints, especially where local file access is expected.
7. Brief developers that address-bar text can be spoofed and that links should be verified through independent navigation rather than displayed URLs.

Source originale

Hackread