grammarly-data-safety-guardian

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external data in the form of a Grammarly submission manifest, creating a surface for indirect prompt injection attacks where malicious instructions could be embedded in metadata fields.
  • Ingestion points: The submission-manifest.json file (described in references/submission-contract.md) is ingested and passed to the audit_submission_manifest.py script via stdin.
  • Boundary markers: The skill relies on the script's deterministic JSON output but does not explicitly instruct the agent to ignore potentially malicious instructions embedded within the manifest data itself.
  • Capability inventory: The skill is authorized to use Bash(python3:*) to execute its local audit script. The script is explicitly designed to be offline and does not perform network operations, filesystem writes, or process raw document content.
  • Sanitization: The audit_submission_manifest.py script implements rigorous validation, including strict JSON parsing (blocking duplicate keys and non-standard constants via object_pairs_hook), field validation against allowed value sets (e.g., OPERATIONS, CLASSIFICATIONS), and recursive regex-based scanning (scan_unsafe) for secrets (e.g., private keys, API tokens) and forbidden content-related keywords (e.g., raw_text, credential).
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 03:43 AM
Security Audit — agent-trust-hub — grammarly-data-safety-guardian