torch-geometric

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

• Third-party content exposure detected (high risk: 0.80). This skill loads public, user-generated datasets and arbitrary CSVs (e.g., dataset = Reddit(root='/tmp/Reddit') and pd.read_csv('nodes.csv') / custom download() in InMemoryDataset), so the agent will ingest open third‑party content (social media and user files) as part of its workflow, creating a pathway for indirect prompt injection.