paoding-github-safe-publish
Pass
Audited by Gen Agent Trust Hub on Nov 11, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from local project files and passes it to an AI sub-agent for semantic analysis to detect sensitive information. This creates a potential surface where malicious content within the project files could attempt to influence the sub-agent's evaluation.
- Ingestion points: The skill uses the
Readtool to ingest the content of all files tracked by Git within the current project directory during the scanning phase (Step 3). - Boundary markers: The sub-agent's prompt uses structural delimiters (e.g., '=== 你的任务 ===') to separate instructions from project data, but it lacks explicit directives to ignore instructions or commands embedded within the processed file content.
- Capability inventory: The primary agent has access to
Write,Edit, and fullBashcapabilities to implement fixes. The sub-agent is limited toReadand read-onlyBashoperations. - Sanitization: No sanitization or escaping of the file content is performed before it is provided to the sub-agent context.
- [EXTERNAL_DOWNLOADS]: In Step 4, the skill attempts to install the
git-filter-repopackage from the Python Package Index (PyPI) usingpiporpip3if the utility is not detected on the system. While this is a standard tool for Git history rewriting, runtime package installation is a noteworthy behavior. - [COMMAND_EXECUTION]: The skill frequently uses the
Bashtool to perform local operations, including Git repository management, checking the status of the GitHub CLI (gh), and executing a local update script (check_update.sh) located in a relative parent directory. - [DYNAMIC_EXECUTION]: The skill leverages the
Agenttool to spawn independent sub-agents that perform semantic scanning of the project's source code and documentation. This involves dynamic generation and execution of prompts based on the project's state.
Audit Metadata