Latent Adversarial Detection: Adaptive Probing of LLM Activations for Multi-Turn Attack Detection — ThinkLLM