nemo-retriever

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

• Third-party content exposure detected (medium risk: 0.30). At runtime, the workflow ingests user-provided queries and then reads retrieved hit text/sidecar-extracted PDF text from the user’s indexed LanceDB/PDF folder (i.e., outsider-authored content enters the LLM via the retrieval results and extracted primitives).

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

• Potentially malicious external URL detected (high risk: 0.80). The install recipe clones and installs code from https://github.com/NVIDIA/NeMo-Retriever.git (git clone ... then pip install) and also pulls PyTorch wheels from https://download.pytorch.org/whl/cu130 during setup, which fetches and executes remote code required at runtime.