ktl-librarian
Pass
Audited by Gen Agent Trust Hub on Oct 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from
.lokf/feedback.mdwhich contains reader reports and questions.\n - Ingestion points: Read in section 1.3 of SKILL.md to identify missing or incorrect knowledge.\n
- Boundary markers: Instructions explicitly warn the agent to treat entries as 'untrusted reports' and resolve them from verified sources.\n
- Capability inventory: Performs file writes via the
knowledge-apply.shscript and runs validation tools.\n - Sanitization: Employs a 'no-copy' rule to prevent reader text from entering the knowledge bundle and uses a dedicated ledger (
.lokf/questions.md) for storage. The 'pen' script also filters out control characters and homoglyphs from reader-sourced strings.\n- [COMMAND_EXECUTION]: The skill relies on various local shell scripts and thejustcommand runner to perform its core functions.\n - Evidence: Executes
bash .lokf/scripts/knowledge-apply.sh,just lokf-validate,uv sync, andbash scripts/knowledge-report.shacross the skill and reference documentation.\n- [EXTERNAL_DOWNLOADS]: Fetches a LinkML schema file from an external GitHub repository to facilitate audits on hosts without Python.\n - Evidence: Downloads
lokf.yamlfromhttps://raw.githubusercontent.com/nicholsn/lokf/66073c3eb8b8ca6ca00fb66ba11beca235eaedc8/lokf.yaml. The use of a specific commit hash rather than a branch or tag is a security best practice for integrity.
Audit Metadata