Fine-tuning open AI models on small, targeted biological datasets can restore dangerous dual-use capabilities excluded during initial pre-training