llm-wiki-implementation-deep-dive
Pass
Audited by Gen Agent Trust Hub on Aug 11, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: The skill includes explicit safety gates and instructions to treat retrieved content as data rather than instructions, which effectively mitigates prompt injection risks. It also warns against direct-write MCP tools and cloud parsers for sensitive data.
- [DATA_EXFILTRATION]: No evidence of data exfiltration or hardcoded credentials was found. The reference documentation specifically advocates for the use of secret scanning (e.g., gitleaks) and PII redaction tools (e.g., Microsoft Presidio) within the ingestion and export pipelines.
- [REMOTE_CODE_EXECUTION]: While the skill references numerous external tools and repositories for LLM-Wiki implementations, it does not provide commands for execution or unverifiable installations. All remote references are for informational and architectural comparison purposes.
- [INDIRECT_PROMPT_INJECTION]: The skill provides an extensive threat model (File:
references/docs/19-security-threat-model.md) that explicitly identifies indirect prompt injection as a primary risk vector. It provides detailed evidence chains for mitigation, including ingestion manifests, parser sandboxing, and human-in-the-loop review gates. - [COMMAND_EXECUTION]: No shell commands or scripts are included in the skill instructions. The skill utilizes a 'pointer pattern' to refer the agent to documentation rather than executing local operations.
Audit Metadata