CerBERTus: a three-headed BERT for LM jailbreaks
Sophos AI has released CerBERTus, a BERT-based model designed to detect jailbreaks in large language models, presented at BSides Las Vegas by Adarsh Kyadige on Monday, August 3.
The model uses a three-headed architecture that applies multi-task supervision to improve both robustness and interpretability compared to standard single-head BERT classifiers. Kyadige’s talk covers the training data, inference strategy, and when multi-task supervision actually helps over a single-head setup.
The second talk: anomaly detection for long-tail threats
On Wednesday, August 5, Ben Gelman, Tamas Nyiri, and Tibor Kristóf Lányi present their anomaly detection research, following on from work first shown at Black Hat USA 2025.
The three are data scientists on the Sophos AI team, and the talk targets a specific gap: supervised classifiers in cybersecurity are typically trained on data that doesn’t capture the long tail of real-world threats. This detector is built for Sophos’ own detection systems, where the long tail is where the misses hurt.
Where it fits in the Sophos stack
CerBERTus and the anomaly detector are part of Sophos AI Defense, the layer that sits alongside Sophos Endpoint and Sophos Fusion in the company’s product stack. Sophos Fusion is the AI-native defense system that currently protects more than 625,000 organizations worldwide, and it is the same system that just got OpenAI’s frontier models via a newly announced partnership.
The OpenAI integration brings frontier model capability directly into the channel and managed service providers through Sophos Fusion, extending the system’s existing coverage of endpoint, network, email, cloud, identity, and security operations. Sophos also joined OpenAI’s Daybreak Cyber Partner Program for managed security services, which is the first time the program has been extended to MSSPs. The three pieces—security from AI, with AI, and for AI—map cleanly to what Sophos is shipping: Endpoint blocks the exploit techniques, the agentic SOC handles detection and response, and AI Defense governs the AI itself.
Key points
- CerBERTus is a three-headed BERT model designed for LLM jailbreak detection.
- Presented by Adarsh Kyadige on Monday, August 3, 2026 at BSides Las Vegas.
- The model uses multi-task supervision to improve robustness and interpretability.
- Sophos AI team is presenting two research talks at BSides LV, with a second anomaly detection talk on August 5.
- Sophos will publish in-depth articles on both projects after the conference.
Further reading & sources
- Three-headed dogs and long tails: Sophos at BSides LV | SOPHOS — Sophos
- When AI doesn’t know the target is real | SOPHOS
- Sophos Ranked #1 Overall in Endpoint, XDR, MDR, and Firewall in G2 Fall 2026 Reports | SOPHOS
- AI shrinks cyberattack timelines as attackers target AI identities, Sophos report finds
- AI compresses cyberattack timelines, targets ungoverned AI identities - IT-Online
- Sophos Working with OpenAI on security from AI, with AI, and for AI | SOPHOS