agent-builder
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill provides instructions to download the official Claude Code installation script from the trusted claude.ai domain.
- [REMOTE_CODE_EXECUTION]: The installation process involves piping a remote shell script from a trusted official source into the bash shell.
- [CREDENTIALS_UNSAFE]: The skill references and accesses the ~/.claude/ directory, which is a sensitive location used by Claude Code to store authentication and session configuration.
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates building agents that process untrusted data while possessing significant capabilities. 1. Ingestion points: External code and pull request data ingested via CLI tools and file reading. 2. Boundary markers: Uses system prompt configuration but lacks explicit implementation of input-ignore delimiters for external data. 3. Capability inventory: Includes full shell access, file modification, and custom MCP tools. 4. Sanitization: No explicit content filtering or escaping for external data is provided in the examples.
- [COMMAND_EXECUTION]: The skill generates code and commands that execute shell operations through the Claude Code CLI headless mode.
- [DYNAMIC_EXECUTION]: Provides templates for generating and running Python scripts that dynamically load and execute agent loops and custom tools.
Audit Metadata