academic-pptx

Pass

Audited by Gen Agent Trust Hub on Jun 3, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill processes untrusted research papers, documents, and web content to plan and generate presentations, creating an indirect prompt injection vulnerability surface.
  • Ingestion points: scripts/source_to_md/pdf_to_md.py, scripts/source_to_md/doc_to_md.py, scripts/source_to_md/excel_to_md.py, and scripts/source_to_md/web_to_md.py.
  • Boundary markers: Absent. No explicit delimiters or 'ignore embedded instructions' warnings were found in the prompts that process document content.
  • Capability inventory: Shell command execution via subprocess.run() in administrative and conversion scripts, file system write access to project directories, and network access to AI model providers.
  • Sanitization: No sanitization or filtering of external document content was detected before it is used for slide drafting.
  • [COMMAND_EXECUTION]: Shell commands are executed via the subprocess module for core tasks such as document conversion and repository updates.
  • doc_to_md.py: Invokes the pandoc executable for converting various document formats into Markdown.
  • project_manager.py: Runs system commands for project initialization, structural validation, and SVG viewBox checks.
  • update_repo.py: A maintenance script that utilizes shell execution for tool updates (likely involving git or pip).
  • scripts/image_gen.py and scripts/image_search.py: Utilize dynamic module loading via __import__ and importlib.import_module to load backend and search providers from internal subdirectories.
  • [DATA_EXFILTRATION]: The skill performs network operations to communicate with external services and fetch web content.
  • AI Backends: Multiple scripts in scripts/image_backends/ and scripts/tts_backends/ communicate with well-known providers like OpenAI, Google Gemini, Anthropic, and ElevenLabs.
  • web_to_md.py: Downloads content from user-provided URLs using requests and curl_cffi (for TLS fingerprint impersonation to bypass bot detection).
  • image_search.py: Connects to external image providers including Openverse, Wikimedia, Pexels, and Pixabay to search for and download assets.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 3, 2026, 08:50 AM
Security Audit — agent-trust-hub — academic-pptx