grammarly-document-evaluator
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document content (e.g., .docx, .txt) which constitutes an indirect injection surface. The ingestion point is the read_document function in scripts/grammarly_api.py. The skill uses a multi-stage approval process requiring an operator to provide an exact SHA-256 digest and a validated manifest, creating a clear human-in-the-loop boundary. The skill includes manifest auditing in scripts/audit_submission_manifest.py to prevent metadata poisoning and credential leakage within policy files.
- [DATA_EXFILTRATION]: The skill performs network operations to transmit document content to Grammarly’s document processing infrastructure. It communicates with api.grammarly.com and AWS S3 (*.s3.amazonaws.com) using established API contracts. The implementation prevents SSRF and DNS rebinding by validating upload hostnames, resolving them to global IP addresses, and pinning the HTTPS connection to those specific addresses in scripts/grammarly_api.py.
- [COMMAND_EXECUTION]: The skill invokes local Python scripts to manage the document evaluation lifecycle. These scripts are composed entirely of Python standard library calls and do not spawn arbitrary shell subprocesses or execute untrusted code.
Audit Metadata