taxonomy-builder
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external paper metadata files to build taxonomies, representing a potential but highly restricted injection surface.\n
- Ingestion points: Reads
papers/core_set.csv,papers/papers_dedup.jsonl,queries.md, andGOAL.mdfrom the workspace.\n - Boundary markers: The script does not use specific delimiters but processes text through structural parsers (CSV/JSONL).\n
- Capability inventory: Limited to writing to
outline/taxonomy.ymland creating a validation marker file; no network or subprocess capabilities are present.\n - Sanitization: Input strings are normalized to lowercase and tokenized using regular expressions to filter for specific concept keywords.\n- [COMMAND_EXECUTION]: The skill relies on a local Python script (
scripts/run.py) to execute its logic.\n - The script uses standard Python libraries and a local
toolingmodule for data processing.\n - It includes logic to dynamically locate the repository root for module imports, which is an expected behavior for internal workspace automation tools.
Audit Metadata