Model Risk

Dolphin3-Cyber-8B, an 8-billion-parameter model fine-tuned from an already-"abliterated" (refusal-removed) base and marketed for offensive-security work with "zero refusals," has passed 44,000 downloads on Hugging Face and runs on an 8GB laptop GPU. What abliteration actually does, why its own disclaimer isn’t an enforced control, and the governance checklist before anyone on your team runs it.