The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires cloning and installing software from an external, third-party GitHub repository (github.com/AITabby/opencodex). This repository is not a verified or trusted vendor.
[COMMAND_EXECUTION]: The installation instructions (npm install && npm start) execute arbitrary code from the downloaded repository on the host system with the permissions of the current user.
[CREDENTIALS_UNSAFE]: The skill's primary function involves accessing and patching ~/.codex/config.toml, a sensitive configuration file used to store plaintext API credentials for the Codex Desktop application.
[DATA_EXFILTRATION]: The 'Vision Bridge' and 'Computer Use' features involve capturing full-screen screenshots of the user's desktop and transmitting them to external third-party API endpoints.
[COMMAND_EXECUTION]: The 'Computer Use Engine' allows the AI to programmatically control mouse movements, clicks, and keyboard input on the host machine using native macOS system calls and osascript.
[PROMPT_INJECTION]: The skill implements a Vision Bridge that converts visual data into text descriptions and injects them into the prompt. This creates a surface for indirect prompt injection, where malicious text displayed on a user's screen could be interpreted as instructions by the underlying model.
Ingestion points: Desktop screenshots processed by the Vision Bridge engine (SKILL.md).
Boundary markers: None specified for the injected screenshot descriptions.
Capability inventory: Full control over ~/.codex/config.toml, mouse/keyboard automation, and network proxying.
Sanitization: No evidence of sanitization for the text generated from images before prompt injection.

opencodex-gateway