@joshfpocock: Comment "CRACKED" and I'll send you all three, plus what each one costs you. One of the best known AI jailbreakers on the internet just posted an uncensored build of Qwen 3.8 27B. Zero refusals across 842 harmful prompts, by their own count. And it isn't one person. Three versions of this same model went up this month under three different names: Uncensored, ABLITERATED, OBLITERATED. Same procedure, different intensities. Here's the part nobody is quoting, and it's published on their own model card. MMLU, the standard test for how smart a model is, fell from 87.4 to 81.4. Six points, gone. A gentler version of the same procedure on the same base model measures about half a point. So the harder you strip the refusals out, the more of the model goes with them. That's the trade, and almost nobody is publishing that curve. How it works: refusal lives in roughly one direction inside the weights. Delete that direction and the model stops saying no. That comes from a peer reviewed NeurIPS paper on AI safety. The part that should get your attention: the tool that does this has 22,000 GitHub stars and has been run on roughly 3,000 models. Every new open weight model gets a version of this within days. #ai #opensource #llm #localai #aisecurity

Josh Pocock

Open In TikTok:

Region: CA

Thursday 20 August 2026 21:17:17 GMT