grammarly-enterprise-rbac
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill handles user-defined role mappings and probe text, which serves as a potential surface for indirect prompt injection.
- Ingestion points: Processes role mappings and probe text during the staging and verification phase in SKILL.md.
- Boundary markers: The instructions include clear boundaries, requiring the use of opaque IDs and forbidding the inclusion of real membership data or document text in logs or tickets.
- Capability inventory: Utilizes fetch for API interactions and Read, Write, and Grep tools for configuration management.
- Sanitization: Explicitly mandates logging only aggregate outcomes and actor metadata while stripping sensitive fields like emails and membership data.
- [DATA_EXFILTRATION]: The provided code snippets demonstrate network interactions with external API endpoints.
- Network operations: The skill uses fetch calls to interact with Grammarly's enterprise API, which is a well-known service.
- Evidence: Network requests are used for authorized administrative tasks like checking permissions and assigning team roles, using variables for credentials rather than hardcoded secrets.
Audit Metadata