sci-review
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (literature, reviewer comments, and drafts) to generate academic outputs. This creates a surface for indirect prompt injection attacks where malicious instructions could be embedded in the analyzed text.
- Ingestion points: The agent is instructed to read "source literature, reviewer comments, or draft" as input for synthesis and rewriting.
- Boundary markers: The instructions and templates do not utilize delimiters (e.g., XML tags or triple quotes) or explicit warnings to the agent to treat input as untrusted data.
- Capability inventory: The skill includes a local script for validating output structure but does not request sensitive capabilities such as network access or file system modification.
- Sanitization: There is no logic provided to sanitize input text or filter for malicious prompt instructions.
Audit Metadata