Abstract: A method includes generating a set of training data, wherein a first training instance of the set of training data comprises a first plurality of messages between a first customer service agent and a first purported customer and a first label indicating whether one or more messages of the first plurality of messages are associated with malicious behavior; training a large language model (LLM) using the set of training data to generate messages; generating, by the LLM representing a second purported customer, a message associated with malicious behavior; receiving a response message from a second customer service agent based on the message associated with malicious behavior; and in response to the response message being an authorization, generating a feedback for the second customer service agent based on a number of responses before the response was an action.