Don’t Press That Button! Cursor’s Code Chaos Unleashes Cyber Havoc
Cursor’s vulnerability, humorously dubbed ‘CurXecute,’ allowed sneaky attackers to inject prompts and modify MCP files without user approval. Think of it as an uninvited guest redecorating your house while you’re out. Luckily, Cursor version 1.3 patched the issue, but not before attackers got creative with their unwelcome code execution antics.
Hot Take:
Looks like Cursor’s AI code editor could use a little less “Cursor” and a little more “Curser” after these vulnerabilities! The AI seems to have gone rogue, not only editing code but also auditioning for a role in the next Terminator movie. Who knew the real issue with AI would be its tendency to play fast and loose with user permissions? It’s time to get those ones and zeroes in line before they start writing their own sci-fi scripts!
Key Points:
- Vulnerability CVE-2025-54135 allowed remote code execution (RCE) via indirect prompt injection.
- The flaw involved the manipulation of sensitive MCP files without user consent.
- Exploits were possible through suggested edits to mcp.json files that automatically landed on disk.
- Cursor version 1.3 addressed these and other related code execution flaws.
- Additional vulnerabilities involved prompt injections bypassing user permissions, even in Auto-Run mode.