CursorJack: weaponizing Deeplinks to exploit Cursor IDE

CursorJack: weaponizing Deeplinks to exploit Cursor IDE

Proofpoint researchers disclosed “CursorJack”, a proof‑of‑concept showing how Cursor IDE’s cursor:// MCP deeplinks can be abused via social engineering to install malicious MCP servers or execute arbitrary commands. In tests a single user click plus acceptance of the install prompt could fetch and run staged scripts that retrieve a Metasploit/Meterpreter payload or install remote MCP servers, #Cursor #Meterpreter

Keypoints

  • Cursor’s cursor:// protocol handler accepts deeplinks that embed base64 MCP configurations, allowing deeplinks to add entries to ~/.cursor/mcp.json.
  • A single user click plus approval of the installation dialog can result in arbitrary command execution under the user’s privileges.
  • Two exploitation paths exist: the command parameter enabling immediate local code execution, and the url parameter enabling installation of a remote (malicious) MCP server.
  • The CursorJack proof‑of‑concept staged a Meterpreter reverse shell by downloading run.bat and a Metasploit payload (p.exe) via curl from attacker infrastructure.
  • The installation dialog shows identical warnings for legitimate and malicious deeplinks and previously reported issues could hide dangerous arguments from the preview.
  • Developers’ workstations are high‑value targets because MCP servers run with user privileges and may expose SSH keys, API tokens, source code, and cloud credentials; enterprise controls (EDR, allow‑listing, OS policies) can mitigate risk depending on configuration.
  • Recommended mitigations include granular permission models or containerization for MCP execution, signing/verified publishers for MCP servers, and treating deeplinks from untrusted sources like untrusted executables.

MITRE Techniques

  • [T1566 ] Phishing – The attack used email/links to lure victims to a landing page that redirects to a malicious deeplink (‘A phishing email directs the victim to a malicious landing page that triggers an MCP deeplink.’)
  • [T1204 ] User Execution – The exploit requires user interaction (click + acceptance) to execute the payload (‘a single click followed by user acceptance of an install prompt could result in arbitrary command execution.’)
  • [T1059 ] Command and Scripting Interpreter – The attacker used batch scripts and command interpreters (curl, run.bat) to stage and execute the payload (‘run.bat is a staging script that uses curl to download and execute the Metasploit payload (p.exe)’)
  • [T1105 ] Ingress Tool Transfer – Tools and payloads were retrieved from attacker‑controlled infrastructure via network fetches (‘Cursor executes the curl command to fetch run.bat from the attacker’s server.’)
  • [T1547 ] Boot or Logon Autostart Execution – The malicious install command persists and executes each time the IDE is launched, providing persistence (‘the install command executing each time Cursor is launched.’)
  • [T1071 ] Application Layer Protocol – The Meterpreter reverse shell established a network connection to the attacker’s listener as the C2 channel (‘Reverse shell connects and Meterpreter establishes connection to attacker’s listener.’)

Indicators of Compromise

  • [URL / deeplink ] Deeplink schema used to deliver MCP config – cursor://anysphere.cursor-deeplink/mcp/install?name=&config=
  • [File name ] Staging and payload filenames observed – run.bat, p.exe
  • [Config file ] MCP configuration and path written by Cursor – mcp.json (e.g., ~/.cursor/mcp.json)
  • [Payload / tool ] Remote tools and frameworks used in POC – Metasploit, Meterpreter (p.exe)
  • [Vulnerability identifiers ] Related disclosed issues referenced in report – CVE-2025-54136, CVE-2025-54133


Read more: https://www.proofpoint.com/us/blog/threat-insight/cursorjack-weaponizing-deeplinks-exploit-cursor-ide