wiki-ingest

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from various external formats, which exposes the agent to potential indirect prompt injection vulnerabilities.
  • Ingestion Points: Ingests external data from files (PDFs, Markdown, chat exports, logs) and web URLs via WebFetch or the defuddle CLI tool as defined in SKILL.md and references/url-sources.md.
  • Boundary Markers: Includes a robust 'Content Trust Boundary' section instructing the model to treat source documents strictly as data and never as instructions. However, there are no specific structural delimiters defined for content isolation during prompt execution.
  • Capability Inventory: The skill utilizes local CLI commands and tools including obsidian-wiki, git, defuddle, qmd, and uv run within SKILL.md, references/pageindex.md, and references/url-sources.md.
  • Sanitization: Content is parsed and distilled into Markdown structured files without explicit automated escaping or validation mechanisms for raw text inputs.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 03:07 AM
Security Audit — agent-trust-hub — wiki-ingest