ktl-sidecar

Pass

Audited by Gen Agent Trust Hub on Oct 22, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONOBFUSCATION
Full Analysis
  • [COMMAND_EXECUTION]: The knowledge-ingest.py script provides a 'command' ingestion mode that executes local scripts located within the repository (e.g., .lokf/ingest/schema-dump.sh). This is used to capture dynamic system data. The script verifies that these commands are tracked by git, ensuring they are subject to repository review processes.
  • [EXTERNAL_DOWNLOADS]: The skill fetches data from external sources including Atlassian Confluence, GitHub, and user-defined HTTP/HTTPS endpoints via the knowledge-ingest.sh script. These downloads mirror external documentation into the repository's .lokf/mirror/ directory. The ingestion manifest explicitly forbids embedding credentials in URLs and uses environment variables for authentication.
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests content from external web sources. This is mitigated through several layers:
  • Ingestion points: External data enters through knowledge-ingest.py into the .lokf/mirror/ path.
  • Boundary markers: The system uses YAML frontmatter and specific Markdown delimiters to isolate fetched content. Agent instructions in ktl-librarian explicitly guide the LLM to treat this content as data rather than instructions.
  • Capability inventory: Powerful capabilities like subprocess.run and network access are present in the core ingestion and application scripts.
  • Sanitization: Content is cleaned of control characters, and the mirrored files are intended for review in a pull request before they are used by the librarian agent.
  • [DYNAMIC_EXECUTION]: The skill utilizes uv to manage and execute its Python environment and dependencies (lokf, PyYAML). It also facilitates the execution of user-configured AI agent CLIs (such as GitHub Copilot CLI or Claude Code) using a wrapper script (knowledge-librarian.sh) that strictly enforces output paths and git state restoration.
  • [OBFUSCATION]: Scripts like knowledge-ingest.py and knowledge-feedback.sh contain logic to handle or filter non-printable Unicode characters. Specifically, knowledge-ingest.py explicitly allows Zero-Width Non-Joiner (U+200C) and Zero-Width Joiner (U+200D) characters to support correct rendering of Persian, Arabic, and Indic scripts in fetched content.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 22, 2026, 07:18 PM