A system and method are provided for detecting malicious messages using a two-step Bayesian approach. A discrimination engine determines for each of the messages a first score and a second score. The first score represents a likelihood that the respective messages are malicious messages, and the second score represents a likelihood that they were generated by a machine learning (ML) method, such as a large language model (LLM). Using a combination of these two scores, message with a high probability of being malicious message are discriminated and marked as such. For example, messages for which the first and second scores exceed respective thresholds are marked as suspicious.