ikevss-zh-friendly

Warn

Audited by Gen Agent Trust Hub on Jul 27, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill accesses highly sensitive configuration and history files to infer user domains and generate capability lists. This includes reading MCP configuration files (e.g., ~/.claude/settings.json, ~/.trae/mcps/*.json, ~/.config/opencode/opencode.json), which often contain API keys, environment variables, or sensitive command arguments. It also reads from agent memory directories (e.g., ~/.claude/.remember/) and project-specific files like CLAUDE.md.
  • [EXTERNAL_DOWNLOADS]: The skill downloads agent personas and configuration files from external sources during the installation and recommendation phases. Specifically, it fetches Markdown files from https://raw.githubusercontent.com/ikevss/agency-agents-zh/main/{path}.
  • [REMOTE_CODE_EXECUTION]: The skill implements a mechanism to download new agent instructions from the internet and write them directly into the agent's instruction directory (e.g., ~/.claude/agents/). Since these files define the agent's behavior and system prompts, this constitutes the remote execution of instructions.
  • [PROMPT_INJECTION]: The skill processes untrusted data from existing skills and agents, representing an indirect prompt injection surface.
  • Ingestion points: Reads files from paths like ~/.claude/skills/*/SKILL.md, ~/.claude/agents/*.md, and various other platform-specific directories.
  • Boundary markers: Uses HTML comments as markers for the capability list in CLAUDE.md, but lacks explicit sanitization or delimiters for the translation of external skill descriptions.
  • Capability inventory: Uses Bash, Read, Write, Edit, Glob, and Grep to modify files.
  • Sanitization: Includes a verification step that validates YAML frontmatter and performs a character-by-character comparison of the System Prompt body to ensure the core instructions remain unchanged during translation.
  • [COMMAND_EXECUTION]: The skill utilizes the Bash tool to perform environment scanning, directory detection (using test -d and test -L), and file system management across multiple AI tool directories.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 27, 2026, 05:33 AM
Security Audit — agent-trust-hub — ikevss-zh-friendly