UNST33R

#abliteration

7 posts

#security-event (8)#abliteration (7)#red-teaming (6)#vulnerability-analysis (6)#open-models (5)#tutorial (5)#model-selection (5)#agent-security (5)#tools (4)#supply-chain (3)#php (2)#ransomware (1)#llm-infra (1)#ai-agents (1)#ai-red-teaming (1)#disclosure-policy (1)#triage (1)#attack-surface (1)#data-report (1)#threat-intel (1)#inference (1)#llama.cpp (1)#policy (1)#open-weights (1)#rce (1)#announcement (1)#deepseek (1)#benchmarks (1)#tiers (1)#model-safety (1)#primer (1)

Abliteration goes one-click: OBLITERATUS and what industrialized refusal removal changes

A 9k-star toolkit turns refusal-direction research into a one-command pipeline with hosted Spaces and crowdsourced telemetry. The technique genealogy, the method menu, and the supply-chain consequences.

Supply signals: why providers are asking for non-abliterated models, and how to choose between abliterated, uncensored, and base

A hosted provider polled users for non-abliterated models; Moonshot ships capability-reduced weights; HF listings vanish. Three supply-side signals, one decision framework for reduced-refusal model selection.

Operation n0n: what the first criminal use of abliterated models actually shows

SOCRadar's STRU unit documented an exposed n0n ransomware workspace where an opencode agent ran uncensored and abliterated models across 154 model entries. The full breakdown: $7,216 in LLM spend, zero CVEs, credential-only access, and what it changes for authorized red-team work.

The abliterated takedown wave: what disappeared from Hugging Face, why, and what users are doing about it

MiMo-UNCENSORED, penclaw-GLM-abliterated, huihui Kimi abliterated GGUFs and more vanished from Hugging Face in October 2026. A fact-checked breakdown of the removals, the backup movement, and the supply-side outlook.

The abliterated model ecosystem, by the numbers (October 2026)

A snapshot of the abliterated ecosystem on Hugging Face: 100 checkpoints, 8.7M downloads, and why Qwen dominates the base-model split. Data pulled October 22, 2026.

Inference-time abliteration: steering refusals away without touching weights

A llama.cpp fork that abliterates any model at load time sparked fresh community discussion. How runtime refusal steering compares to baked-in weight edits for red-team harnesses.

What is abliteration?

How abliteration removes refusal directions from open-weight LLMs via orthogonal weight projection, its tradeoffs for tool calling, and how it differs from fine-tuning or jailbreaks.