FAR.AI releases TamperBench, finding that safety training in open-weight LLMs can be stripped in a few hundred fine-tuning steps · Digg