1-minute read of a 6 min report

Cursor Flaws Let Prompt Injection Overwrite Its Sandbox and Run Code

By

1-minute read

What a defender needs to know

Cursor 3.0 fixed two path-handling flaws that allowed an AI coding agent to write outside its project workspace. Cato AI Labs named the pair DuneSlide. CVE-2026-50548 let the agent choose the working directory used to build the sandbox’s writable area. CVE-2026-50549 allowed a write through an in-workspace symlink when path canonicalization failed.

Cato demonstrated how attacker-controlled content in an MCP response or poisoned search result could steer an agent toward either bug. The researchers then overwrote Cursor’s sandbox helper so later commands could run without the restriction. No public evidence currently ties the flaws to a real intrusion.

Update Cursor Desktop to version 3.0 or later. Then inventory coding agents that can execute commands or write files automatically, including personally managed installations used for company work. Check whether an agent can influence the parameters that define its own workspace, and confirm that failed path or policy checks block the action. Limit access to powerful cloud, repository and package credentials in agent-enabled development sessions.

Primary sources