Three-Headed Dogs and Long Tails: Sophos at BSides LV
Sophos, Friday, July 31st, 2026
Sophos AI researchers preview BSides Las Vegas talks on a three-headed jailbreak detector and LLM-assisted data labeling.
Sophos AI researchers will present two talks at BSides Las Vegas. The first introduces CerBERTus, a three-headed BERT model that detects LLM jailbreak attempts by separately analyzing harmful intent, underlying goal, and presentation style rather than treating detection as binary classification.
The second details an automated pipeline that combines anomaly detection with large language models to identify and label unusual cybersecurity data that traditional supervised classifiers miss.
The two projects address complementary gaps: prompt injection risk on one side, and false positives and negatives in threat detection on the other.
Sophos argues both approaches improve robustness and interpretability across different data types.