fact-check
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted markdown documents provided by the user for fact-checking.
- Ingestion points: Document content is read by
scripts/init.sh,scripts/build-claims.py, andscripts/validate-claims.sh. This content is then passed to LLMs during Phase 1 (Extraction) and Phase 5 (Deep Verify). - Boundary markers: The system prompt in
prompts/extract-claims.mduses explicit 'Rules' and a 'Verbatim Copy Rule' to constrain LLM behavior, but lacks strict structural delimiters for the input document itself. - Capability inventory: The skill has access to web search tools (
anysearch), the ability to spawn subagents for parallel processing (subagenttool), and performs network requests viaurllibandcurlinscripts/rule-engine.sh. - Sanitization: The skill uses
scripts/validate-claims.shto perform deterministic checks (JSON syntax, schema, and exact substring matching against the source) which mitigates some risks of the LLM 'hallucinating' or being diverted by instructions embedded within the document text.
Audit Metadata