Generative-AI models often face security threats such as prompt injections and data exfiltration. Cybersecurity firms are fighting fire with fire — using AI to secure LLMs — but there are costs.