A computing system may obtain text corresponding to a conversation between an outside caller and an agent. The computing system may obtain data associated with the conversation that may be used to determine whether the outside caller is attempting malicious activity or not. The obtained text and data may be provided to a machine learning model to generate a probability score indicative of whether the outside caller is attempting to obtain unauthorized access or attempting other malicious activity. Based on determining that the probability score satisfies a threshold, the computing system may modify a permission (e.g., an API permission) of the agent. The computing system may deny the request to perform the action and may cause display of an indication that the request was successful.