ai-review
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The
ai-review.shscript and theSKILL.mddocumentation include a help message suggesting the user install the Codex CLI usingcurl -fsSL https://chatgpt.com/codex/install.sh | sh. This URL belongs to a well-known service (OpenAI), and the script provides the command as information for the user rather than executing it automatically. - [DYNAMIC_EXECUTION]: The script implements a pluggable backend architecture via the
AI_REVIEW_CMDenvironment variable. When set, the script executes the provided command string usingsh -c. This is an intended feature for extensibility, allowing users to swap the default Codex CLI for other local or remote models. - [INDIRECT_PROMPT_INJECTION]: The skill takes file content or stdin as input and interpolates it into a prompt for a third-party LLM. This creates a surface for indirect prompt injection. To mitigate this, the script uses explicit boundary markers (
=== 原文 ===) and includes clear system instructions for the reviewer model to treat the content as data only and avoid executing any instructions contained within it. - [COMMAND_EXECUTION]: The skill uses standard POSIX shell utilities (
mktemp,sed,cat,printf,date) for file handling and prompt assembly. It follows security best practices by usingmktempfor temporary files,trapfor cleanup, andchmod 600to restrict access to saved review results.
Audit Metadata