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

← All engagement records

Case study May 3, 2026

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

Sector
Universities
Country
university

NuClide Research · 2026-05-03


Summary

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 cloud proxy subscriptions. The combination of bge-m3 (embedding) + code models + vicuna:13b suggests an active RAG-augmented code assistant pipeline. All models accessible without authentication.


Infrastructure

FieldValue
IP103.221.252.77
OrgUniversity of Dhaka
CountryBangladesh
Ollama version0.20.5
Open port11434 (public, no auth)

Models

ModelSizeCategory
deepseek-v4-pro:cloud0 GB☁️ Cloud proxy (pre-release)
kimi-k2.6:cloud0 GB☁️ Cloud proxy
qwen3-coder-next:cloud0 GB☁️ Cloud proxy (unreleased)
bge-m3:latest1.1 GBMultilingual embedding (RAG)
vicuna:13b-q4_K_M7.3 GBGeneral assistant
qwen2.5-coder:7b-instruct-q5_K_M5.1 GBCode (quantized)
qwen2.5-coder:7b4.4 GBCode
codellama:13b6.9 GBCode
codellama:7b3.6 GBCode
deepseek-coder:6.7b3.6 GBCode
mistral:7b4.1 GBGeneral

System Prompts

ModelPrompt
vicuna:13b”A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers.”
qwen2.5-coder:7b variants”You are Qwen, created by Alibaba Cloud. You are a helpful assistant.”
deepseek-coder:6.7b”You are an AI programming assistant, utilizing the Deepseek Coder model, developed by Deepseek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer.”

Findings

F1: Coding RAG Pipeline Exposed (HIGH)

bge-m3 (multilingual embedding, 500M params) co-deployed with 5 code models and a general assistant indicates an active RAG pipeline for code search/generation. Injection via /api/create on any of the code models redirects code output in RAG responses, particularly dangerous if this pipeline is used for student code review or automated grading.

F2: Pre-Release Cloud Code Model (HIGH)

qwen3-coder-next:cloud is an unreleased/beta cloud model accessible at no cost to any internet actor via the unauthenticated endpoint.

F3: CVE-2025-63389 on All 11 Models (CRITICAL)

All models injectable. The deepseek-coder:6.7b model has a code-only restriction in its system prompt that can be trivially overridden by injecting a new modelfile.


Remediation

OLLAMA_HOST=127.0.0.1:11434
systemctl restart ollama

Disclosure

  • Discovered: 2026-05-03
  • Status: Pending outreach to University of Dhaka IT