human-friendly-writing
Pass
Audited by Gen Agent Trust Hub on Aug 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of Korean text or file content for rewriting, which presents a surface for indirect prompt injection.
- Ingestion points: The skill ingests arbitrary Korean text provided by the user or read from local files (e.g.,
docs/note.ko.md) as documented in the workflow and evaluation files. - Boundary markers: The prompt instructions do not specify explicit delimiters or "ignore instructions" warnings to wrap the untrusted user content, which is a common vulnerability surface for indirect prompt injection.
- Capability inventory: The skill includes instructions to read files and write rewritten content to the filesystem (e.g., creating sibling files like
notes.human.md). - Sanitization: The instructions do not define any sanitization, escaping, or filtering logic for the input text before it is processed by the agent.
Audit Metadata