Most recent
navigate open esc close Corpus index built 2026-06-07 23:58 UTC

§ THE STACK / GATEWAY LAYER

RAG Frameworks

LangChain, LlamaIndex, Dify, retrieval pipelines

Routes the request, attaches retrieved context, mediates between user and model.

What it is

Retrieval-Augmented Generation is how an LLM gets access to documents it wasn’t trained on: your company wiki, last week’s invoices, a PDF of your medical history. A RAG pipeline chains a document loader, an embedder, a vector store, a retriever, and the LLM call. Frameworks that package this into one runtime: Dify (the most polished, Chinese-origin), Flowise (visual builder on top of LangChain), Haystack (Deepset’s enterprise stack), Quivr, Verba. The pipeline is what turns a model into a product.

What goes wrong

Most RAG deployments are research artefacts that grew into prototypes that grew into production. Dify ships with admin@admin.com / password as the seed account; a fresh Flowise install exposes the canvas and every workflow’s embedded API keys; Haystack’s REST API is unauthenticated by default and its /query endpoint will dutifully retrieve and return any document the embedder has indexed. The corpus exposed this way ranges from public PDFs all the way to attorney-client communications, internal sales decks, and patient records.

How we test

We probe each framework’s signature endpoints: Dify’s /console/api/setup for the seed-account state, Flowise’s /api/v1/chatflows for the workflow catalogue, Haystack’s /search for the indexed corpus reach. When the retriever is reachable, we issue a single low-volume query (e.g. “summary”) to confirm the corpus contains real content, capture the document titles and sources from the response, and stop. Title metadata is enough to attribute the operator and characterise the data class without reading the documents themselves.

Receipts

Research

Every survey, case study, and disclosure we've published that touches this layer of the stack. Counts on the cells above tally these directly.

Cross-cloud surveys

7
Survey Jun 6, 2026

Dify Population Survey — 939 Config-Disclosure, 9 Open Auth Findings

Dify is an open-source LLM application development platform (drag-and-drop workflow builder, RAG pipelines, agent orchestration). 2,289 Shodan-indexed instances on http.title:"Dify".

Read →
Survey May 31, 2026

RAG Framework Servers Population Survey — Cat-07 (2026-05-31)

First population survey of the RAG-framework-server category. 16 platforms in the 2026-05-27 pre-assessment intel (data/platform-intel/rag-frameworks-osint-2026-05-27.md); 15 dorks run this session. T…

Read →
Survey May 28, 2026

RAG Stragglers: LightRAG, RAGFlow, DocsGPT, Ragapp Population Survey

Four RAG platforms were left unfinished from prior survey runs: LightRAG, RAGFlow, DocsGPT, and Ragapp. This pass closes them out with a full Shodan harvest, verification, and arsenal run.

Read →
Survey May 15, 2026

RAG Framework Servers: Population-Scale Survey (2026-05-15)

<!-- ksat-tag:auto-generated:start --> ## DCWF KSAT coverage

Read →
Survey May 3, 2026

MinIO + Dify on Public Cloud: Auth Posture Survey

Two parallel sweeps:

Read →
Survey May 1, 2026

Embedding Services: Cross-Cloud Survey (2026-05)

Embedding servers are the vector-conversion layer between raw text and vector databases. They ingest documents or queries and return dense float vectors; without them, RAG pipelines and semantic searc…

Read →
Survey May 1, 2026

RAG Framework Servers: Cross-Cloud Survey (2026-05)

RAG (Retrieval-Augmented Generation) framework servers sit between vector databases and LLM clients. They orchestrate the document-ingestion → chunking → embedding → retrieval → context-injection pipe…

Read →

Field cases

19
Case May 15, 2026

23.239.19.219: Exposed LlamaIndex Chat with Broken Backend, Multi-Tenant SNI Co-Tenancy

23.239.19.219. Linode US datacenter (Akamai AS), 23.239.0.0/19, rDNS 23-239-19-219.ip.linodeusercontent.com. Linode shared-allocation, neighbor at .217 is harperdbcloud.com. No AS63949 honeypot salt m…

Read →
Case May 3, 2026

University of Dhaka: Coding Cluster, 3 Cloud Proxies, Embedding Pipeline

University of Dhaka (AS137359) exposes an Ollama instance focused on software development AI tooling: multiple code-specialized models, a high-quality multilingual embedding model (bge-m3), and three…

Read →
Case May 3, 2026

China Telecom Tianjin: 46-Node Multi-Tenant Ollama Cluster

China Telecom's Beijing-Tianjin-Hebei Big Data Industry Park (Tianjin, AS141679) hosts at least 46 cloud VM instances running Ollama on port 11434 without authentication. All discovered through a Shod…

Read →
Case May 3, 2026

Agricultural University of Athens: 142GB Qwen3-235B MoE, Dual-Embedding RAG

afa4pc19.aua.gr (143.233.187.19), Agricultural University of Athens (Γεωπονικό Πανεπιστήμιο Αθηνών, AUA), runs Ollama v0.18.2 with a 5-model stack anchored by qwen3:235b-a22b-instruct-2507-q4KM, the Q…

Read →
Case May 3, 2026

Institut Teknologi Bandung (ITB): 22 Models, Custom Indonesian Education AI

Institut Teknologi Bandung's LSKK (Laboratorium Sistem Komputer dan Kecerdasan Buatan, Computer Systems and AI Lab, Electrical Engineering) exposes Ollama at lskk-20.ee.itb.ac.id (167.205.66.20) with…

Read →
Case May 2, 2026

Government AI Infrastructure Exposures

Unauthenticated Ollama instances discovered on government networks. Identified via hostname TLD filtering (.gov, .go.id, .gov.br, .gov.tw, .mil, etc.).

Read →
Case May 2, 2026

Indonesia Government Cluster: 5-Node Survey, 2 Account Takeovers

Five Indonesian government Ollama nodes confirmed live across .go.id infrastructure. Two provincial government nodes have live Ollama Connect account takeover URLs. The cluster spans national, provinc…

Read →
Case May 2, 2026

DINAS KOMINFO PROV. JAWA TENGAH: Account Takeover, RAG Pipeline

The Central Java Province Communications and Information Technology Department (Dinas Kominfo Prov. Jawa Tengah) exposes an Ollama node at sijoli-11-245-107.jatengprov.go.id (103.107.245.11) on the In…

Read →
Case May 2, 2026

"No. 18 Institute of Jingdong HQ": 26-Node Cluster, China Unicom

A 26-node Ollama cluster on China Unicom's 111.228.0.0/16 range, all registered to org eleven street,No. 18 Institute of Jingdong headquarters. The org name reads as a Chinese physical address (Jingdo…

Read →
Case May 2, 2026

Kyungpook National University: 3-Node Cluster, Multimodal AI

Kyungpook National University (KNU, Daegu, South Korea) exposes three Ollama nodes on the public internet. Together the nodes span vision-language models (qwen3-vl:32b, llava), a custom community quan…

Read →
Case May 2, 2026

California Institute of Technology (Caltech): GPT-OSS 120B, RAG Pipeline

A Caltech node (yertle.caltech.edu, 131.215.141.46) exposes Ollama with 6 models including gpt-oss:120b (OpenAI's 120B open model, 65.4GB) and a RAG pipeline stack (two embedding models). The hostname…

Read →
Case May 1, 2026

Chinese Primary School: Cloud Proxy Subscriptions + Credential Leak

An Experimental Primary School in China (Shodan org: "Experimental Primary School") is running Ollama with three cloud proxy subscriptions, DeepSeek V4 Pro, Devstral-2 (123B), and MiniMax M2.7, alongs…

Read →
Case May 1, 2026

University of Newcastle, Australia: DeepSeek Cloud Proxy + RAG Pipeline

University of Newcastle (Australia, Callaghan campus) server with deepseek-v4-pro:cloud cloud proxy subscription and mxbai-embed-large:latest embedding model indicating an active RAG pipeline. Raw Oll…

Read →
Case May 1, 2026

Brno University of Technology: Abliterated Gemma + Bulgarian GPT + RAG Pipeline

Brno University of Technology (VUT Brno), Czech Republic, is running Ollama on a Faculty of Mechanical Engineering server with an abliterated Gemma 3 model (safety fine-tuning removed), two variants o…

Read →
Case May 1, 2026

Technical University of Crete + NTUA: Unauthenticated Ollama, MiniMax Cloud, 235.7B Model

- Username: arian - SSH pubkey: ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIASZr/fN5P73o/WF6vT/owMFz3ftTeBlzOpEFpS2QStP - Cloud proxy: minimax-m2.7:cloud (MiniMax API subscription)

Read →
Case May 1, 2026

University of Crete Medical Center: Dual-Embedding RAG Pipeline

The University of Crete Medical Center (centaur.med.uoc.gr) is running Ollama with a sophisticated dual-embedding RAG pipeline, both mxbai-embed-large and nomic-embed-text are deployed alongside large…

Read →
Case May 1, 2026

Fu Jen Catholic University: Medical Public Health GPU Server, 75GB + 60GB Local Models

Fu Jen Catholic University's Medical and Public Health department (user220.medph.fju.edu.tw) has an Ollama instance exposed on port 11434 with 8 models totalling over 200GB of local inference capacity…

Read →
Case May 1, 2026

Rochester Institute of Technology: 4-Node Cluster, DGX with 18 Cloud Subscriptions, Student Machine with Abliterated Models

Rochester Institute of Technology (RIT) has four externally-accessible Ollama nodes on campus, including an NVIDIA DGX research server with 18 cloud proxy subscriptions (same subscription portfolio as…

Read →
Case May 1, 2026

SUNY Buffalo: Unauthenticated Ollama + Cloud Proxy Quota Hijack Confirmed

State University of New York at Buffalo research compute node running 26 Ollama models including gemma4:31b-cloud, a cloud proxy model. Cloud proxy inference confirmed live, 200 OK response at operato…

Read →