Horizon3 has set up the Adversarial Intelligence Frontier Lab. The lab researches closed-loop systems that detect and close attack paths in production environments. The company is using attack data from NodeZero tests at its customers as the foundation. A researcher from Purdue University heads the lab.
Horizon3 says the lab takes a different starting point from much AI security research, which often relies on sensor telemetry or tests against digital twins. The new lab starts with verified end-to-end attack paths that have already been successfully executed against live systems. According to the company, the research focuses on keeping offensive and defensive systems safe, observable and controllable in production environments.
Research lead from Purdue
Z. Berkay Celik will be the lab’s first director. He is an associate professor of computer science at Purdue University and co-director of the Purdue Security Laboratory. His research focuses on the security of complex software and systems, with an emphasis on detecting vulnerabilities, analysing attacks, and designing and evaluating defence methods for autonomous and agentic systems. Celik has previously led research projects with partners from academia, industry and government.
Horizon3 researchers have already published academic papers. One of these is Closing the Sim-to-Real Gap: An Evaluation Framework for Autonomous Cyber Defense Configuration of Commercial EDR, an evaluation framework for the autonomous configuration of commercial EDR solutions. Other publications cover deception techniques against AI attackers and an LLM architecture for autonomous cyber defence.
Four research areas
The lab is starting out with four research areas. The first concerns AI-assisted attacks. The lab is investigating how AI models accelerate vulnerability discovery and how attack tools and coordinated agent swarms work and scale. It is also looking at how defenders can detect and disrupt such attacks.
The second area concerns adaptive defence. The lab is researching how AI attacks bypass security measures. It is also examining how offensive and defensive agents can together speed up the cycle of hacking, fixing and verifying. Adaptive deception and the secure sharing of findings between environments also fall under this area.
The third area focuses on controllable autonomy. This covers architectures and evaluation methods for autonomous agents, with attention to scope limitations, interruption and recovery, and repeatable measurements of effectiveness and safety. The fourth area concerns the cyber resilience of critical infrastructure and coordination between operational domains. It also includes the skills and training that future security professionals will need.
Horizon3 also announced that the lab will accelerate the training of customer-specific cyber world models. According to the company, these models form the shared context and long-term memory for red and blue agents. Horizon3 expects this to enable the agents to coordinate better during attack and defence cycles.
Training data from production environments
The research draws on data from Horizon3’s more than 7,000 customers, including banks, government agencies, manufacturers and hospitals. Hundreds of thousands of NodeZero tests on production systems generate labelled training data. Horizon3 estimates that it collects 30 billion tokens of anonymised and aggregated data every day.
This data covers every attack carried out, whether successful or not. It also records the responses of defensive tools and the configuration of services behind the firewall. The company says this provides a more detailed picture of the interaction between attacker and defender than datasets based on sensor telemetry, synthetic environments or digital twins.
Snehal Antani, CEO and co-founder of Horizon3, says the company has spent years collecting the necessary attack and defence data. According to Antani, the next step is to use verified attack paths. With these, the company aims to improve collaboration between offensive and defensive agents in production environments.
Critical infrastructure
The lab will also oversee Horizon3’s Cyber Resilience Centers of Excellence. These centres focus on critical infrastructure. Horizon3 expects the centres to translate the research findings into practical applications, specifically in environments where an outage or breach could have national or economic consequences. To do so, the centres use validated attack path information and customer-specific cyber models.
The full report is available at Closing the Sim-to-Real Gap: An Evaluation Framework for Autonomous Cyber Defense Configuration of Commercial EDR.
This article was originally published in Dutch on MSP Business. Read the original →
