codex-cli

Warn

Audited by Gen Agent Trust Hub on Mar 13, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Enables the execution of arbitrary shell commands through the codex exec subcommand and an interactive shortcut in the TUI.
  • [COMMAND_EXECUTION]: Documents the --dangerously-bypass-approvals-and-sandbox (or --yolo) flag, which explicitly removes all filesystem and network restrictions, granting the AI agent unrestricted system access.
  • [EXTERNAL_DOWNLOADS]: Recommends installing the @openai/codex package and running @modelcontextprotocol/server-github via npx. While the openai scope is trusted, this specific package and the referenced 'GPT-5' models are not recognized official components.
  • [PROMPT_INJECTION]: The skill acts as a surface for indirect prompt injection because it reads and processes project files without implementing security boundaries or sanitization.
  • Ingestion points: Local project directory files and configuration files.
  • Boundary markers: Not implemented.
  • Capability inventory: Arbitrary shell command execution, filesystem modification, and package installation.
  • Sanitization: No sanitization or validation of external codebase content is mentioned.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 13, 2026, 05:39 AM