wikieval
Pass
Audited by Gen Agent Trust Hub on Jul 20, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill identifies an indirect prompt injection surface as it processes external content from both golden files and candidate outputs.
- Ingestion points: Reads markdown files located in
llmwiki/wiki/sources/evals/and candidate JSON files provided via the--outputsparameter. - Boundary markers: The system uses YAML frontmatter in its golden files to structure data fields, effectively separating input and expected results from other content.
- Capability inventory: The engine executes a local script (
harness/scripts/wikieval.py) and has file system write access to update regression baselines. - Sanitization: It employs a deterministic 'tier 1' assertion cascade (regex, JSON, and SQL-like checks) performed in pure code to validate outputs before any optional model evaluation.
- [COMMAND_EXECUTION]: The skill provides instructions to execute the local Python script
harness/scripts/wikieval.pyfor performing evaluation suites and checking regression gates.
Audit Metadata