Talking Nonsense: Probing Large Language Models' Understanding of Adversarial Gibberish Inputs