nature-writing
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted user-provided content, such as research notes, claims, and drafts, which serves as a surface for indirect prompt injection.
- Ingestion points: User research notes, claims, figures, and results are ingested through the
SKILL.mdrouting logic and processed by fragments likestatic/fragments/language/zh-to-en.mdandstatic/fragments/language/en.md. - Boundary markers: The instructions lack explicit delimiters or warnings for the agent to ignore instructions that might be embedded within user-supplied research data.
- Capability inventory: The skill is capable of generating manuscript prose and filling LaTeX templates (
templates/submission/). It does not demonstrate high-privilege capabilities like network exfiltration or command execution. - Sanitization: There are no instructions for sanitizing, escaping, or filtering the content of the user-provided notes before they are interpolated into the drafting prompt.
Audit Metadata