Exploring the Adversarial Capabilities of Large Language Models