scientific-slides
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSCREDENTIALS_UNSAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves taking user-provided text descriptions and incorporating them into prompts for an external AI model.
- Ingestion points: The prompt parameter in generate_slide_image.py and generate_schematic.py accepts untrusted data from the user.
- Boundary markers: Documentation in SKILL.md defines how the agent should structure these prompts, providing a form of instructional boundary.
- Capability inventory: The skill can write files to the local file system and perform network requests via the requests library.
- Sanitization: The Python scripts do not perform explicit sanitization or filtering of the user-provided prompt before sending it to the external API.
- [COMMAND_EXECUTION]: Multiple scripts use the subprocess module to run helper scripts or system utilities like pdflatex.
- Evidence: scripts/generate_schematic.py, scripts/generate_slide_image.py, and scripts/validate_presentation.py call subprocess.run().
- Security: These calls are executed safely without shell=True, and arguments are passed as lists, preventing shell command injection.
- [EXTERNAL_DOWNLOADS]: The skill makes network connections to an external service to perform its core functions.
- Evidence: generate_schematic_ai.py and generate_slide_image_ai.py use the requests library to connect to https://openrouter.ai/api/v1.
- Context: This is the intended behavior for an AI-powered generation tool using well-known model hosting infrastructure.
- [CREDENTIALS_UNSAFE]: The skill manages the OpenRouter API key using environment variables or local configuration files.
- Evidence: The scripts use python-dotenv to look for keys in a .env file.
- Context: This is a standard and recommended practice for managing secrets in local automation scripts and is not considered a vulnerability.
Audit Metadata