nature-paper2ppt
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content (scientific papers, PDFs, and links) and has the capability to write files and access the network. This creates a surface for indirect prompt injection where adversarial content within a paper could influence the agent's output.\n
- Ingestion points: User-provided scientific papers in various formats (PDF, DOI, arXiv, text) are loaded into the agent's context through tools like PyMuPDF.\n
- Boundary markers: The skill uses logical structure and quality rules to guide the agent, but lacks explicit delimiters or strict instruction-guarding wrappers for the ingested paper content.\n
- Capability inventory: The agent can write
.pptxfiles usingpython-pptxand execute a local Python audit script (scripts/audit_pptx_quality.py). It also performs network requests to resolve paper links.\n - Sanitization: The skill includes checks for 'AI-template phrases' and structural quality audits, which serve as quality gates but do not specifically sanitize adversarial instructions embedded in paper text.\n- [COMMAND_EXECUTION]: The skill instructs the agent to execute a provided local script (
scripts/audit_pptx_quality.py) to perform structural and quality audits on the generated PowerPoint files. This script uses the Python standard library and operates locally on the generated output.
Audit Metadata