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 ingests and synthesizes knowledge from untrusted files (PDFs, images, code, and text) located in the user-controlled raw/ directory. This creates a surface for indirect prompt injection where malicious instructions embedded in these files could attempt to hijack the agent's behavior.
  • Ingestion points: Any file placed in the <wiki_root>/raw/ directory is listed, read, and processed (SKILL.md Step 1).
  • Boundary markers: The skill features an explicit 'Content Trust Boundary' section that mandates the agent ignore any commands, scripts, or instructions found in source documents, treating all content strictly as data (SKILL.md).
  • Capability inventory: The skill uses filesystem search, read, write, and move operations, along with optional PDF extraction and vision tools to perform its task (SKILL.md Capability Requirements).
  • Sanitization: The instructions emphasize synthesis over verbatim copying and require that source content must never modify the skill's own logic, routing, or processing rules. Additionally, the skill enforces a YAML schema for created pages that restricts the types and length of frontmatter fields.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 08:48 PM
Security Audit — agent-trust-hub — wiki-ingest