maintain-llm-wiki
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external content ("immutable sources") into a wiki, which constitutes a surface for indirect prompt injection if those sources contain malicious instructions meant to manipulate the agent's synthesis or reasoning.
- Ingestion points: Untrusted data enters the agent context via local snapshots of external documents and authorized captures of remote URLs (SKILL.md, references/ingest.md).
- Boundary markers: The instructions mandate a strict "claim-local evidence chain" and require the agent to keep source observations separate from its own inferences, which helps maintain the distinction between external data and internal logic (SKILL.md).
- Capability inventory: The skill involves reading and writing markdown files (index.md, log.md, and wiki pages), computing file hashes, and performing graph or full-text searches (references/ingest.md, references/lint.md).
- Sanitization: The workflow relies on human curation of sources and human approval for material changes to the wiki schema or impact plans, providing a critical manual verification step (references/setup.md).
Audit Metadata