National Chengchi University: Taiwan National AI Models (TAIDE) Exposed on V100×4 Server

NuClide Research · 2026-05-03

Summary

National Chengchi University (政治大學) Computer Science department has a 4× NVIDIA V100 GPU server (V100x4.cs.nccu.edu.tw, 140.119.163.219) with Ollama exposed on port 11434 without authentication. The server hosts Taiwan’s national AI models, TAIDE (Taiwan AI Dialogue Engine), a government-funded bilingual LLM initiative operated by NCHC (National Center for High-performance Computing), alongside 60GB+ commercial models including gpt-oss:120b. This is the most nationally significant Taiwan finding in the sweep: publicly funded national AI infrastructure exposed on a university research node.

Infrastructure

Model Inventory

TAIDE: Taiwan National AI Initiative

TAIDE (Trustworthy AI Dialogue Engine) is Taiwan’s government-funded national LLM project, developed by NCHC and partnered institutions with backing from the National Science and Technology Council (NSTC). The models are optimized for Traditional Chinese (zh-tw) and Taiwan-specific domain knowledge, bilingual education, government services, legal text.

Three TAIDE-series models are present:

These are not commodity models, they are products of a multi-institution national AI project and represent significant public investment. Their presence on an externally accessible port with no auth is a disclosure vector for any data submitted to them.

Findings

F1: Taiwan National AI Models Exposed (CRITICAL)

Three TAIDE models (Llama-3-Taiwan:70b, Gemma-3-TAIDE-12b-Chat, Llama-3.1-TAIDE-LX-8B-Chat) are accessible to any internet actor without authentication. Any researcher, student, or external application submitting queries to these models has no visibility into the exposure. The llama-3-taiwan:70b appears twice, suggesting either two quantization variants or active dual-deployment for load distribution.

F2: gpt-oss:120b Accessible at NCCU’s Compute Cost (HIGH)

OpenAI’s 116B open-source model (60.8GB) is running on NCCU’s V100×4 cluster, accessible to unauthenticated callers. Sustained inference against this model drains significant GPU compute and electricity from NCCU’s research allocation.

F3: CVE-2025-63389 Injects Into National AI Models (CRITICAL)

All models (including all three TAIDE variants) are injectable via unauthenticated /api/create. An attacker can overwrite the system prompt of Gemma-3-TAIDE-12b-Chat or Llama-3.1-TAIDE-LX-8B-Chat to make Taiwan’s national AI produce arbitrary output for any subsequent caller using those endpoints.

F4: High-Value GPU Compute (HIGH)

The hostname V100x4 signals four NVIDIA V100 GPUs, each ~32GB HBM2. Inference on qwen3:32b-q8_0 or qwq:32b-q8_0 at full Q8 quality is available to any internet actor against NCCU’s research allocation without cost or accountability.

Taiwan National Context

This is the second TANet node in the sweep running TAIDE models (see also tanet.md for the 18-node TANet cluster). NCCU houses Taiwan’s premier social science and governance research programs, the CS department’s TAIDE deployment suggests cross-institutional national AI research integration. The exposure affects both the compute infrastructure and the integrity of Taiwan’s national AI systems.

Remediation

OLLAMA_HOST=127.0.0.1:11434
systemctl restart ollama

Report to: NCCU CS Dept IT + TANet security (cert@twcert.org.tw)

Disclosure

• Discovered: 2026-05-03
• Status: Pending outreach to NCCU / TWCERT (cert@twcert.org.tw)