Large language models (LLMs) remain vulnerable to generating unsafe outputs, even after undergoing alignment training. To mitigate this risk, a real-time monitoring system can be implemented to detect and alert when the model's output may compromise safety. This system utilizes a verifier signal from an external model, which is then thresholded to trigger an alarm decision, with the threshold calibrated through risk control. Experimental results demonstrate the effectiveness of this approach in maintaining safety standards. The development of LLMs, such as those by ARM, introduces new capabilities and risks, with security implications often overshadowed by hype1. As LLMs become increasingly prevalent, the importance of online safety monitoring grows, highlighting the need for robust risk management strategies to prevent potential harm. The integration of such monitoring systems is crucial for ensuring the safe deployment of LLMs, and their absence can have significant consequences for users and organizations relying on these models.