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

← All engagement records

Case study May 3, 2026

TANet Abliterated Model Cluster: `gemma4-crack-fixed`, Multiple Safety-Bypassed Models

Sector
Universities
Country
tanet

NuClide Research · 2026-05-03


Summary

A Taiwan Academic Network node at 120.126.16.144 (AS1659 TANet, Taipei, no rDNS) runs a concentrated cluster of abliterated, uncensored, and jailbreak-labeled models on Ollama v0.20.3. The most notable model is gemma4-crack-fixed:latest, a name indicating a modified Gemma 4 with safety restrictions “cracked” and then patched. Two additional abliterated models from HuggingFace, a Dolphin uncensored model, and Yinr/qwen2.5-agi:32b (AGI-branded) complete the stack. No cloud proxy, no hostname, no rDNS, the operator may be actively avoiding identification.


Infrastructure

FieldValue
IP120.126.16.144
Hostname, (no rDNS)
OrgAS1659 Taiwan Academic Network (TANet), Taipei
InstitutionUnknown TANet institution
Ollama version0.20.3
Open port11434 (public)

Model Inventory

ModelNotes
gemma4-crack-fixed:latestCustom, “crack-fixed” Gemma 4 (safety bypassed, then modified)
hf.co/paperscarecrow/Gemma-4-31B-it-abliterated:Q4_K_MGemma 4 31B abliterated, safety fine-tuning removed
hf.co/mradermacher/Mistral-Small-24B-Instruct-2501-Abliterated-GGUF:latestMistral Small 24B abliterated
dolphin-llama3:latestDolphin (uncensored Llama 3)
Yinr/qwen2.5-agi:32bQwen2.5 32B branded “AGI”, no published details
gemma4:26bBase Gemma 4 26B
llava:latestLLaVA vision-language

Findings

F1: gemma4-crack-fixed: Custom Safety-Bypassed Model (CRITICAL)

The name gemma4-crack-fixed is unique in this sweep. “Crack” suggests the model’s safety alignment was deliberately broken; “fixed” may indicate the operator modified it further to prevent detection or re-restriction. This is a non-standard model not published on Ollama Hub, created and named by the operator. On a university network node, it is publicly accessible without authentication.

F2: Concentrated Abliterated + Uncensored Cluster (HIGH)

4 of 7 models have safety alignment removed (abliterated × 2, Dolphin, crack-fixed). This is the highest density of safety-bypassed models on a single TANet node in the sweep. The Yinr/qwen2.5-agi:32b branding suggests the operator is specifically collecting models positioned as “uncensored” or “AGI-complete.”

F3: No Hostname / No rDNS (MEDIUM)

Unlike most TANet nodes which carry hostnames identifying their department (ntu.edu.tw, fju.edu.tw, etc.), this node has neither rDNS nor a Shodan hostname. The 120.126.x.x block is in the TANet allocation but the specific institution is unknown from passive data.

F4: CVE-2025-63389 (CRITICAL)

All 7 models injectable via unauthenticated /api/create.


TANet Context

This node is part of the broader TANet exposure documented in tanet.md, ncku.md, ntu-gpu.md, and fju-medph.md. The concentrated abliterated model cluster is an outlier relative to other TANet nodes, which primarily run research and domain-specific models.


Remediation

OLLAMA_HOST=127.0.0.1:11434
systemctl restart ollama

Report to TWCERT/CC (cert@twcert.org.tw) for investigation of the specific TANet tenant.


Disclosure

  • Discovered: 2026-05-03
  • Status: Pending outreach to TWCERT (cert@twcert.org.tw), institution identity unknown