use-chatgpt-by-applescript

Warn

Audited by Gen Agent Trust Hub on Oct 9, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONOBFUSCATION
Full Analysis
  • [DYNAMIC_EXECUTION]: The chatgpt-research-runner.mjs script performs runtime compilation of the bundled chatgpt_status.c file using clang. It generates a binary in /tmp/ or ~/.local/bin/ to interface with the macOS ApplicationServices framework for UI monitoring. Additionally, it dynamically generates and executes AppleScript blocks via osascript to control the ChatGPT application's GUI.
  • [COMMAND_EXECUTION]: The main runner script utilizes execSync to invoke a variety of system tools including osascript, pbcopy, ssh, scp, and clang. These commands are used to manage clipboards, execute remote instructions, and compile code on the fly.
  • [PRIVILEGE_ESCALATION]: The skill requires the user to manually grant 'Accessibility' permissions to system processes like sshd-keygen-wrapper or terminal applications. This is a high-privilege permission on macOS that allows the skill to inspect and interact with all elements in the graphical user interface, effectively allowing it to 'see' and 'type' as the user.
  • [OBFUSCATION]: The skill uses Base64 encoding to transport prompt payloads through the clipboard and over SSH. While documented as a technical solution for shell character escaping, it results in the execution of instructions that are not visible in plain text during transit.
  • [INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its processing of external data.
  • Ingestion points: Untrusted data enters via the --prompt CLI argument or through the targets.json file processed by chatgpt-research-runner.mjs.
  • Boundary markers: The skill prepends plugin URI tags but does not use clear delimiters or safety warnings to isolate user-supplied prompts.
  • Capability inventory: The skill possesses significant capabilities including UI automation (osascript), remote command execution (ssh), runtime compilation (clang), and the ability to scrape UI text (ApplicationServices).
  • Sanitization: There is no evidence of sanitization, filtering, or validation of the input prompt text before it is encoded and delivered to the target application.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 9, 2026, 10:27 PM