grant-planning
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the agent to download scientific paper source archives from the well-known
arxiv.orgrepository (e.g.,https://export.arxiv.org/e-print/). These downloads are targeted at LaTeX source files for literature review and evidence verification within the grant planning process. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content from scientific papers, which creates a potential surface for indirect prompt injection.
- Ingestion points: External LaTeX, bibliography, and metadata files downloaded into the
/Background/papers/directory (SKILL.md). - Boundary markers: The instructions include explicit directives to distinguish between source-backed facts and interpretations, and to verify results directly from source passages rather than relying on summaries.
- Capability inventory: The skill utilizes network access for specific well-known domains and performs file system operations to extract and store archives.
- Sanitization: The skill explicitly mandates the rejection of archive entries containing absolute paths, directory traversal sequences (
..), or links that could write outside the designated directory to protect against path traversal attacks.
Audit Metadata