codexmcp-claude-codex-collaboration

Fail

Audited by Gen Agent Trust Hub on Sep 16, 2026

Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Fetches the installation script for the uv tool from astral.sh. As this is a well-known service for development tools, the download itself is considered standard for the skill's purpose.
  • [REMOTE_CODE_EXECUTION]: The installation guide includes commands that pipe remote scripts directly into the shell for execution (curl ... | sh and irm ... | iex). This pattern is common in developer workflows but bypasses content validation before execution.
  • [REMOTE_CODE_EXECUTION]: The skill installs the codexmcp server directly from a third-party GitHub repository (github.com/GuDaStudio/codexmcp.git) using uvx during the setup process.
  • [PRIVILEGE_ESCALATION]: The Windows installation instruction utilizes the -ExecutionPolicy ByPass flag to execute the remote script, which temporarily overrides PowerShell's security restrictions for that specific command execution.
  • [COMMAND_EXECUTION]: The codex tool allows the agent to execute coding tasks that can involve file system operations. The skill includes built-in safety documentation for sandboxing, defaulting to a read-only mode and providing warnings for high-privilege modes like danger-full-access or yolo.
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect injection as it ingests user prompts and reads files from the working directory to perform coding tasks.
  • Ingestion points: Data enters the context via the PROMPT parameter and files residing in the cd working directory path.
  • Boundary markers: The skill explicitly instructs the agent in its system prompt to treat Codex output as a "unified diff patch" and "logical reference" only, and to verify all changes before implementation.
  • Capability inventory: The skill can perform file writes and execute code through the Codex tool depending on the sandbox parameter configuration.
  • Sanitization: Security relies on the provided sandbox levels and the requirement for manual review of tool outputs as described in the documentation.
Recommendations
  • HIGH: Downloads and executes remote code from: https://astral.sh/uv/install.sh - DO NOT USE without thorough review
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 16, 2026, 03:57 AM