scholar-evaluation

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.75). The skill’s runtime path ingests outsider-authored free text from the Gemini 3.1 Pro Preview review model (external LLM output) via generate_schematic_ai.py -> review_image() -> _make_request() -> content/reasoning which is then used in subsequent prompts (improve_prompt) and thus fed back into the agent’s LLM context; this is not user-authored content.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 1.00). The scripts call the OpenRouter API at runtime (base URL: https://openrouter.ai/api/v1 and https://openrouter.ai/keys for API keys) and use the API's textual critique and image outputs to decide whether to regenerate and to automatically modify prompts, and the tool requires an OPENROUTER_API_KEY, so external content directly controls prompt/instruction flow.