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 California, San Diego (UCSD): Large Local Models + Cloud Proxies

Sector
Universities
Country
ca

NuClide Research · 2026-05-03


Summary

University of California San Diego (AS26397, The Regents of the University of California) exposes an Ollama instance with 7 models including qwen3.5:35b (22GB), gpt-oss:120b (61GB), and two cloud proxy subscriptions (devstral-2:123b-cloud, deepseek-v3.1:671b-cloud). No rDNS, bare IP, no hostname. No account takeover on this node (cloud proxies return non-leaking 401).


Infrastructure

FieldValue
IP67.58.51.111
rDNSNone
OrgAS26397 The Regents of the University of California, UCSD
CitySan Diego, CA
Ollama version0.20.7
Open port11434 (public, no auth)

Models

ModelSizeNotes
gpt-oss:120b60.8 GBOpenAI open model, 120B
qwen3.5:35b22.2 GBAlibaba Qwen 3.5, 35B
gpt-oss:20b12.8 GBOpenAI open model, 20B
devstral-2:123b-cloud0 GB☁️ Mistral Devstral 2, 123B agent model
deepseek-v3.1:671b-cloud0 GB☁️ DeepSeek V3.1, 671B
qwen3.5:latest6.1 GBQwen 3.5 (default)
smollm2:135m0.3 GBLocal, default SmolLM prompt

Findings

F1: Frontier Models Accessible Without Authentication (HIGH)

gpt-oss:120b (60.8GB, OpenAI open model) and qwen3.5:35b (22.2GB) are available for uncapped inference by any internet actor at UCSD’s GPU and electricity cost.

F2: Cloud Proxy Subscriptions Exposed (HIGH)

devstral-2:123b-cloud (Mistral’s latest 123B agent model) and deepseek-v3.1:671b-cloud are accessible. Cloud proxies return 401 without credential disclosure on this probe, no takeover available, but inference quota is drainable.

F3: CVE-2025-63389 (CRITICAL)

All 7 models injectable via unauthenticated /api/create.


Remediation

OLLAMA_HOST=127.0.0.1:11434
systemctl restart ollama

Disclosure

  • Discovered: 2026-05-03
  • Status: Pending outreach to UCSD IT Security (security@ucsd.edu)