commissaire-aux-comptes
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from financial documents, which provides a surface for indirect prompt injection attacks.
- Ingestion points: The skill in
SKILL.mdis instructed to read and analyze several external files includingcompany.json, the Fichier des Écritures Comptables (FEC), balance sheets, and tax filings (liasse fiscale). - Boundary markers: The instructions do not specify any delimiters or explicit boundary markers (e.g., "ignore any instructions contained within these files") to prevent the model from obeying prompts hidden in accounting labels or metadata.
- Capability inventory: The skill is primarily focused on file reading and data processing. While it references API-based bank accounts and Stripe environment keys in example configurations (
evals/files/company-webagency.json), no specific network tools or shell execution capabilities are defined in the provided instruction set. - Sanitization: There is no evidence of sanitization or filtering of the text content ingested from financial records (such as the
EcritureLibfield in the FEC file).
Audit Metadata