codex-advisor
Warn
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: MEDIUMEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill directs users to install the global package
@openai/codexvia npm and a cask namedcodexvia Homebrew. These are not official or standard packages distributed by OpenAI, creating a potential supply chain risk or dependency confusion vector. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from external sources such as git diffs and arbitrary project files, creating an indirect prompt injection vulnerability surface.
- Ingestion points: Untrusted content is ingested from codebase paths via
-C .,git diff, and direct file targets such assrc/auth/login.tsinSKILL.md. - Boundary markers: Absent. The command templates append or pipe raw data directly into the execution prompt without formatting wrappers or isolation tags.
- Capability inventory: The skill triggers command-line execution using the
codex execcommand. - Sanitization: No input validation, character escaping, or filtration mechanisms are defined to sanitize the inputs before passing them to the external CLI.
Audit Metadata