codex
Fail
Audited by Gen Agent Trust Hub on Jul 15, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill's documentation directs users to install a global NPM package named
@openai/codexand to download the skill's instruction file directly from a personal GitHub repository (199-biotechnologies/codex-skill). - [REMOTE_CODE_EXECUTION]: The skill utilizes the agent's bash capabilities to execute an external CLI tool (
codex). Specifically, the 'Delegated Mode' enables a--full-autoflag which, according to the instructions, allows the third-party binary to autonomously explore the codebase, write files, and run arbitrary commands. - [COMMAND_EXECUTION]: Significant portions of the skill are dedicated to constructing and executing complex shell commands that pass potentially sensitive project context to the
codexbinary. - [OTHER]: The skill contains highly deceptive claims, primarily the assertion that it provides access to 'GPT-5.4', a model version that does not currently exist. Additionally, the referenced NPM package
@openai/codexis not an official OpenAI product, suggesting a social engineering attempt to induce the installation of untrusted software with broad system permissions.
Recommendations
- AI detected serious security threats
Audit Metadata