Trojan Activation Attack: Red-Teaming Large Language Models using Steering Vectors for Safety-Alignment