The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input via the prompt argument in scripts/generate_schematic.py. This input is interpolated directly into meta-prompts for external AI models (Gemini) within scripts/generate_schematic_ai.py without boundary markers or sanitization. This surface allows malicious user input to potentially influence the image generation and quality review processes performed by the external models.
Ingestion points: User-provided prompt argument in scripts/generate_schematic.py.
Boundary markers: Absent; input is directly concatenated into prompt templates.
Capability inventory: Spawning sub-processes (subprocess.run), network requests to external APIs (requests), and local file writing.
Sanitization: Absent; no validation or filtering of the input string is performed.
[DYNAMIC_EXECUTION]: The scripts/build_decision_tree.py utility and various LaTeX templates in the assets/ directory generate .tex files by interpolating user-provided text. If this input contains malicious LaTeX macros (e.g., \input for file reading or \write18 for shell execution), these commands are written into the document. Local compilation of these files with tools like pdflatex could lead to unintended code execution or information leakage depending on the user's local environment configuration.

clinical-decision-support