Horizon3 launches AI security lab with 30 billion daily tokens
Horizon3's new lab will train attack-and-defence AI on 30 billion tokens of daily production data from 7,000 customers, led by Purdue professor Dr. Z. Berkay Celik.

Why it matters
- Horizon3 launched the Adversarial Intelligence Frontier Lab on November 16, 2026, led by Purdue's Dr. Z. Berkay Celik.
- The company says it collects 30 billion tokens of anonymised, labelled attack-defence training data daily.
- Horizon3 reports more than 7,000 customers, including banks, federal agencies, manufacturers, and hospitals.
- The lab begins with four research areas, including AI-enabled attacks, attack-and-defence learning loops, safe autonomy, and cyber resilience.
- Cyber Resilience Centres of Excellence will apply the research to critical infrastructure sectors.
Horizon3 says it collects 30 billion tokens of labelled attack-and-defence training data every day — and it has now built a research lab to put that data to work. On November 16, 2026, the company launched the Adversarial Intelligence Frontier Lab, an AI cyber defence research group led by Dr. Z. Berkay Celik, Associate Professor of Computer Science at Purdue University and Co-Director of the Purdue Security Laboratory.
The bet here is unusual. Most AI security projects train on sensor telemetry or on tests run in synthetic environments. Horizon3's lab will instead work from verified attack paths that have already succeeded against live systems, using evidence from production-safe runs of the company's NodeZero attack-graphing tool to find and close attack paths in real environments.
"We've spent years collecting the attack-and-defence data required to research these systems. The next step is to use that verified attack-path evidence to improve how offensive and defensive agents work together in environments that actually matter," said Snehal Antani, Chief Executive Officer and Co-Founder of Horizon3.
The launch lands in the middle of a sector-wide push. Security companies and researchers are moving fast into AI for both offensive and defensive cyber tasks, and the open question is whether systems trained on real production evidence can stay observable and controllable while cutting exposure to genuine attack paths.
Who is running the lab?
Dr. Celik joins as the lab's inaugural director while keeping his Purdue posts. His research record covers the security of complex software and systems: vulnerability discovery, attack analysis, and the design and evaluation of defences for agentic and autonomous systems.
That background maps directly onto the lab's mandate. The whole project is about agents — offensive ones, defensive ones, and how they interact — and Celik's work has centred on making such systems defensible and measurable.
The lab's stated priority is keeping attack-and-defence systems safe, observable, and controllable in production. That framing matters because autonomous hacking tools that can't be watched or rolled back are a liability, not an asset.
What are the four research areas?
The lab starts with four defined strands:
- AI-enabled attacks and adversarial behaviour. This includes studying how frontier AI models may speed up vulnerability discovery, and how emerging attacker tools operate — including coordinated agent swarms.
- Adaptive defence and attack-and-defence learning loops. The work will examine how AI-driven attacks bypass existing security controls, and how offensive and defensive agents can be coordinated through repeated cycles of hacking, fixing, and verification.
- Safe and measurable autonomy. This strand covers the design of autonomous security agents and the methods used to measure them, including scope constraints, interruption and rollback mechanisms, and reproducible measures of effectiveness, defensive outcomes, and safety.
- Cyber resilience and operational readiness. Researchers will look at how AI changes cyber operations and defence activity across operational domains, plus the skills and training future cyber workers will need.
The common thread is closed-loop learning: attack, observe what happens, fix, verify, repeat. Horizon3 wants that loop running on real evidence rather than simulations.
Where does the training data come from?
The data advantage is the core of the pitch. Horizon3 says it has more than 7,000 customers, spanning banks, federal agencies, manufacturers, and hospitals, and that these organisations generate labelled training data from production-safe NodeZero tests.
The daily haul — 30 billion tokens of anonymised, aggregated, and labelled attack-defence data — includes:
- attacker commands
- whether those commands succeeded or failed
- how defensive tools responded
- contextual information on how key services are configured behind the firewall
That gives the lab records drawn from live environments rather than telemetry feeds, synthetic settings, or digital twins. Horizon3 argues this provides a stronger basis for studying real interactions between attackers and defenders than data gathered any other way.
The dataset also feeds a second deliverable: customer-specific cyber world models. These are intended to act as a shared context and long-term memory layer for red and blue agents, so attacking and defending agents can coordinate more effectively across repeated attack-and-defence cycles. In plain terms, the agents get a common picture of the environment they're fighting over.
What about critical infrastructure?
The lab will also oversee Cyber Resilience Centres of Excellence focused on critical infrastructure. These centres will apply the lab's research to essential systems in sectors where downtime or cyber compromise can have national or economic consequences.
The mechanism is translation, not invention. The centres take research findings and turn them into operational improvements using production-validated attack-path intelligence and customer-specific models, with the stated aim of strengthening resilience in environments where service disruption has wider consequences.
This is the part of the programme with the highest stakes. A learning loop that works on a bank's test environment is one thing; the same loop applied to a hospital or a manufacturer carries a different order of risk, which is why the lab's emphasis on scope constraints and rollback mechanisms is not decorative.
What has Horizon3 already published?
The lab builds on prior Horizon3 research into AI attacker behaviour, cyber deception, and autonomous defence. That body of work has covered deception techniques aimed at AI attackers, methods for evaluating autonomous cyber defence settings, and tool-mediated architectures for autonomous cyber defence.
The new lab formalises and extends that line of inquiry under dedicated academic leadership. Hiring a sitting Purdue professor as director signals the company wants peer-grade rigour attached to its data asset, not just a marketing badge.
The takeaway is simple: Horizon3 is betting that verified, production-scale attack evidence — 30 billion tokens a day from 7,000 customers — beats synthetic training environments for building cyber AI that holds up in the real world. Whether the sector's central question, keeping such systems observable and controllable while reducing real attack exposure, gets answered will depend on what this lab and its rivals can demonstrate in live environments.
Source: itbrief.com.au
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
633 articles
Related articles
- Trillium Labs Launches to Do High-Stakes AI Research in the Open
- Frontier AI Labs Score No Better Than C+ on Public Controls
- One Hacker, a Few Thousand Dollars, and an AI Army That Breached John Deere
- One Israeli Startup Sits Behind a String of Rogue AI Disclosures
- DOJ Launches “Super Intelligence” Task Force Over AI Misuse