Verba at 567 title matches: a single-user RAG app behind an external identity gap
Verba is a retrieval-augmented generation application built on Weaviate. It imports documents, creates a knowledge base, and lets a user query that material through a web interface. The project is useful for local research and small deployments, but its design has a security property that should be explicit in any internet-facing inventory: the maintainers describe it as a single-user application without multi-user or role-based access control.
On November 18, 2026, ZoomEye returned 567 matches for title="Verba" and six for app="Verba". The request used sub_type=all, page 1, and page size 1. "Verba" is also a common word and a legal term, so the title count must not be read as 567 live RAG deployments. The six application matches are a narrower product signal, but they still do not prove exploitability or configuration.
The official Verba repository lists deployment through Docker or local Weaviate and describes the frontend at port 8000 and Weaviate at port 8080. It also says the application is designed for single-user use and that there are no plans for multi-user or role-based access in the near future. The FAQ says the FastAPI layer is intended for communication between the frontend and backend, not as a general external API.
The data and the model keys are the asset
A Verba instance can contain imported documents, embeddings, collection metadata, and configuration for model providers. The environment may also include a Weaviate API key, an OpenAI-compatible model key, an Unstructured endpoint, or other service credentials. A user who reaches the interface may be able to query the corpus or change the RAG configuration. A user who reaches the backend may be able to do more.
If the application has no built-in account boundary, adding a reverse proxy login is the minimum control. That login must protect both the frontend and the FastAPI paths; a rule that covers only the HTML page leaves the backend reachable. Network policy should also prevent the container from accepting connections from the public internet. The service is not designed to be a shared multi-tenant portal, so placing it behind a generic login without testing object-level access does not turn it into one.
Credentials need separate treatment. Use a dedicated Weaviate account for Verba, not an admin key for the whole vector database. Give the model provider key a spending limit and a deployment-specific name. Do not store the key in a document collection or in a prompt that users can retrieve. If the instance is exposed, rotate the Weaviate, model, and ingestion credentials before deleting the old deployment.
What the counts mean
The title count is large enough to produce false positives. The application count is small, which is useful for a targeted review but not enough to conclude that only six systems exist. Many deployments may use a custom title, a reverse proxy, or a page that does not expose "Verba" in the indexed title. Some may run entirely on an internal network and never appear in public scan data.
Start with the owned application inventory and compare it with the scan. For each deployment, identify the imported data classification, the model providers, the Weaviate endpoint, and the network path from the internet. Check whether the frontend and backend are both protected by the same access policy. Test whether an unauthenticated request to the FastAPI documentation or a data endpoint returns content.
If Verba is being used as a personal research tool, keep it on localhost or a private development network. If teams need a shared knowledge service, use a product with an explicit multi-user authorization model or put a well-tested application gateway in front of it and document the limitations. A single-user RAG interface is not a safe place to store a mixed collection of confidential documents just because it has a login page from a reverse proxy.
The article count is a reason to look at the architecture. It is not evidence that a specific Verba instance has been compromised.
Query evidence
title="Verba"
sub_type=all, page=1, pagesize=1, observed 2026-11-18T07:42:37+08:00, count=567
Top comments (0)