unglaze
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill instructions focus entirely on linguistic style and information density. It defines specific transformation rules, banned phrases, and formatting defaults to ensure output is scannable and technically accurate.
- [INDIRECT_PROMPT_INJECTION]: The skill identifies an ingestion surface by processing untrusted data from external files, user-pasted blobs, and referenced URLs. However, because the skill's capabilities are strictly limited to text rewriting and it does not possess tools for file system modification, network exfiltration, or code execution, the risk of instruction override via processed data is negligible.
- Ingestion points: Processes content from previous assistant turns, local files, pasted text, and referenced URLs (specified in
SKILL.mdworkflow). - Boundary markers: None explicitly defined in the instructions for segmenting untrusted content.
- Capability inventory: No subprocess calls, network operations, or file-writing tools are requested or used.
- Sanitization: No specific sanitization or escaping logic is implemented for external content.
Audit Metadata