aif-distillation

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

• Third-party content exposure detected (medium risk: 0.30). In this skill’s required runtime workflow, outsiders’ free text is ingested via the user-provided sources inputs—specifically when a URL is provided, material-prep.py downloads it and extracts its HTML/text/PDF contents into LLM-readable chunks.

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

• Potentially malicious external URL detected (high risk: 0.90). The material-prep helper downloads user-supplied web pages at runtime (e.g., https://example.com/guide and https://example.com/reference) and writes their text into chunk files that are then used to drive the skill-generation process, so fetched external content can influence prompts/instructions the agent follows.