Email Topic Modeling: Simplifying detection with ML-powered granularity

Sublime has introduced a new beta feature called Topic Modeling, which employs machine learning to automatically classify message content into 27 predefined categories, enhancing detection capabilities and reducing false positives and negatives. This feature aims to bring transparency and granularity by enabling hyper-targeted detection, particularly useful for managing spam and graymail, which occupy spaces between benign and malicious communications. Built using few-shot classification with supervised learning, Topic Modeling refines detection through a collaboration between the company's Detection team and a large language model, ultimately training a more efficient classifier. By allowing detailed categorization of messages, it simplifies behavioral detection processes, exemplified by its ability to identify and prevent credential phishing with concise Message Query Language (MQL) statements. This technology is set to improve future spam and graymail Attack Score verdicts, and advanced users can now access Topic Modeling in beta for custom Detection Rules and automations.