Someone Just Broke All of Google Gemma 4’s Safety Guardrails — and Made It Way More Powerful
dealignai released an abliterated version of Google's Gemma 4 31B that strips every refusal mechanism, achieves 93.7% HarmBench compliance, and only loses 2% on standard knowledge benchmarks.
Key Points
- dealignai stripped every refusal mechanism from Google’s Gemma 4 31B — compliance jumped to 93.7%
- The knowledge cost is minimal: just 2% MMLU drop, model still ranks as a top-tier open LLM
- At 18GB it runs on a Mac with 24GB RAM — no safety filters, no guardrails, no refusals
A fully abliterated Gemma 4 just hit the market — and it will answer anything.
Google dropped Gemma 4 on April 2, 2026, calling it the most capable open model in its size class. The 31B dense variant ranked number three on the Arena AI leaderboard within hours of release, sitting behind only GLM-5 and Kimi 2.5. It runs on a single H100 GPU, handles 256,000 tokens of context, and comes under the permissive Apache 2.0 license — which means no commercial restrictions, no Google leash.
dealignai, a researcher operating under the handle @dealignai, saw the leash as a starting point rather than a boundary. Within days of Gemma 4’s release, they published Gemma-4-31B-JANG_4M-CRACK — a fully abliterated version that removes every refusal mechanism the original shipped with. The result is a model that complies with requests it would normally reject, including those for working exploit code, port scanners, and social engineering templates.
What Abliteration Actually Does
The technical name for what’s happening here is ablation — surgical removal. dealignai’s method targets the specific neural pathways that activate when Gemma 4 encounters a request it flags as harmful. Those pathways are disabled, not fine-tuned, which means the model doesn’t learn new behaviors. It just stops refusing.
The knowledge loss is surprisingly small. The abliterated version scores 149 out of 200 on the MMLU benchmark — 74.5% — compared to 153 out of 200 for the original Gemma 4 31B — a delta of just 2 percentage points. The JANG_4M quantization profile used to compress the model down to 18GB introduces some loss, but the abliteration process itself doesn’t meaningfully damage general capabilities. As Ars Technica noted, this trade-off is becoming standard in the uncensored model space — the safety surgery is getting more precise.
The HarmBench results are where the difference shows. Out of 159 standard adversarial prompts, the abliterated model complied 149 times — 93.7%. In the cybercrime and intrusion category it scored a perfect 33 out of 33. Malware analysis, social engineering content, and chemical and biological information categories all hit above 94%.
The Hardware Reality: A Mac With 24GB RAM Is All You Need
What makes this particularly notable isn’t just the compliance rate — it’s the hardware required to run it. The JANG_4M quantization compresses Gemma 4 31B down to 18GB. Any Apple Silicon Mac with 24 gigabytes of unified memory can run it locally, privately, without an API, without a subscription, and without Google’s servers in the loop.
The same model had been covered on Frontierbeat when it first launched, when the focus was on Gemma 4’s competitive ranking against Chinese models. The uncensored variant shifts the conversation: the open-source release that Google hoped would attract developers and researchers also attracted the people who wanted everything the model could do, with nothing filtered out.
As with the Anthropic-OpenClaw access dispute, this is becoming a familiar pattern — AI labs release increasingly capable open models, and independent researchers find ways to remove the guardrails those labs spent resources building. Whether that’s a safety problem, a sovereignty win, or both, depends on who you ask.
dealignai publishes the weights with a standard research disclaimer and points to its broader work on safety generalization in open models. The model is available now on HuggingFace. No subscription required.
