deeplearning-ai-agentic-research

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

• Third-party content exposure detected (high risk: 0.95). The workflow uses the Tavily web search tool at runtime (tavily_search_tool → requests.post to https://api.tavily.com/search), then concatenates the returned r['content'] (free-form web text) into research_data, which is later embedded into the LLM prompt in writer_agent via context = "\n\n".join([... f"{source}\n{data}" ...]) and sent as the user message.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill calls external APIs at runtime—notably https://api.tavily.com/search and http://export.arxiv.org/api/query (and Wikipedia via its API)—and injects the fetched content into the model context/messages, so those remote URLs can directly influence prompts and model behavior.