Back Issues/Search Home → Calendar → Archive → RSS → Subscribe → Services → Vendors → Blog → Current Issue → Popular → Newsletters (LinkedIn, Substack)

All issues › Volume 344, Issue 2 › IT Vendor News › Sophos

Your Training Data Is Boring: Anomaly Detection with LLMs

Sophos, Wednesday, November 11th, 2026

Sophos AI researchers describe a pipeline that pairs anomaly detection with LLM labeling to add unusual samples to security training data.

Sophos data scientists Ben Gelman, Tamas Nyiri and Tibor Kristof Lanyi presented research at BSides Las Vegas 2026 on improving training data for cybersecurity classifiers.

They argue that benign and malicious training sets are dominated by common patterns, leaving the long tail of unusual files unlabeled and prone to misclassification in production.

Their automated pipeline uses complementary anomaly detection methods to surface unusual data, then has an LLM classify each anomaly with format-specific prompts.

Two separate prompts, one erring toward benign and one toward malicious, must agree before a label is accepted, which improved precision, and the method was evaluated on command lines, HTML and RTF.

more →  ·  More from Sophos →