Case study · NVIDIA RAG Blueprint · AI infrastructure
NVIDIA's RAG Blueprint shipped an MCP server that read user-supplied file paths with no validation. Any client could read arbitrary files and ingest them into the RAG collection.
The RAG Blueprint is NVIDIA's reference architecture for retrieval-augmented generation. Its MCP server lets agents upload documents into the pipeline, reading each path from disk before forwarding the contents. It bound to 0.0.0.0:8000 with no authentication.
Both upload tools passed file paths straight to open() with no checks. A client could supply ../../../../etc/passwd, have it ingested into the RAG collection, then read it back through the model. A filesystem read laundered into natural-language answers.
Path validation that canonicalises each path with os.path.realpath() and rejects anything outside a configurable base directory.
Tell us what you are shipping and we will scope a focused review. Andesyte can stay with your team as the code changes.