deep-learning

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

• Third-party content exposure detected (high risk: 1.00). The skill explicitly fetches and ingests arbitrary public web and social-media content (e.g., SKILL.md "Material Collection" lists web articles, X/Twitter, Reddit, Medium, LinkedIn via jina-reader, and scripts/orchestrate.sh uses jina-reader/curl and notebooklm source add-research to import URLs and search results), and that untrusted content is then read and used to drive artifact generation and follow-up actions, so it can enable indirect prompt injection.

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

• Potentially malicious external URL detected (high risk: 1.00). The orchestrator fetches external content at runtime via a curl to "https://r.jina.ai/$material" (in scripts/orchestrate.sh) and injects that fetched Markdown into NotebookLM as a text source, which directly supplies model context/instructions—this meets the criteria for a runtime external dependency that can control prompts.