Eset Launches NeoCortex: An AI-Powered Cybersecurity Model Built In-House
Eset, the Slovak cybersecurity firm, has officially launched NeoCortex, an artificial intelligence (AI) model developed using its own security data and hosted on the company’s servers. Rather than opting for a large, generic model, Eset has created a more focused solution tailored to predict the progression of cyberattacks. By keeping client telemetry within its infrastructure, the firm ensures critical decision-making remains with human analysts.
Eset unveiled NeoCortex on October 30, with a formal announcement in France on November 3. This foundational model underpins LiveCortex, the engine that correlates security signals within Eset’s enterprise protection platform, Eset PROTECT. The company describes NeoCortex as a “model of the world,” designed to learn how information systems operate and how events unfold.
This launch fulfills a commitment made back in May at Eset World in Berlin, where the company announced a €40 million investment in AI by 2026, which includes developing foundational models trained specifically on cybersecurity data. At that time, LiveCortex was introduced as a nascent system. On October 1, Eset added AI agent control to its platform for its clients.
A Specialized Model, Trained and Hosted In-House
Trained on historical incidents, NeoCortex reconstructs missing context during security events, synthesizes complex attack patterns, and identifies anomalies within event sequences. According to Eset, LiveCortex processes billions of telemetry events daily, with only 0.00111%—or roughly one in 90,000—escalating to clients as incidents. This filtering aims to reduce alert fatigue for security teams inundated with more signals than they can analyze.
Eset justifies the specialization of NeoCortex based on the analytical demands of cybersecurity. Unlike a generic large language model that strives for plausible responses across a range of topics, security investigations require a consistent tracking of entities, timestamps, indicators of compromise, and process relationships throughout an incident. NeoCortex focuses solely on this specific terminology and does not possess extended conversational capabilities.
The decision to create a specialized model also stems from cost considerations. Internal research indicates that training a model like NeoCortex can be up to 200 times cheaper than developing a large general-purpose model. Eset attributes this efficiency to a cybersecurity-centric training focus, reduced computational needs, and the use of proprietary hardware. However, this figure is based solely on the company’s own assessments, lacking an independent verification protocol.
Eset’s second justification revolves around client data control. Developed and operated on Eset’s internal infrastructure, NeoCortex keeps sensitive client telemetry and model functionality under the company’s supervision. This approach minimizes reliance on third-party AI providers and mitigates exposure to potential commercial or geopolitical shifts affecting these vendors. Juraj Malcho, Eset’s CTO, emphasizes the importance of maintaining cybersecurity intelligence “under our control,” reinforcing human expertise rather than replacing it.
The Human Analyst as the Sole Decision-Maker for Critical Responses
One of Eset’s most ambitious promises with NeoCortex is its predictive capability—estimating the next likely step an attacker might take, enabling defenders to act preemptively. Eset terms this approach “defense J-1,” emphasizing anticipation rather than guaranteeing a 24-hour lead time. The company highlights the week leading up to October 25, referred to as “Black October,” when critical vulnerabilities in PHP, Linux, and GitLab necessitated swift, prioritized responses.
Importantly, the model does not dictate actions automatically, and Eset clarifies that NeoCortex is not the sole decision-making entity. It operates alongside specialized models, deterministic controls, correlation rules, and human expert oversight. NeoCortex functions within constrained environments using read-only tools, operational limits, continuous monitoring, and a shutdown switch. Critical decisions and responses ultimately remain the responsibility of a human analyst.
This emphasis on human intervention mirrors moves by CrowdStrike, which recently introduced a shutdown switch for AI agents that deviate from their intended tasks. Such safeguards are becoming essential in the governance of AI-driven agents, whether for defense tools or assistants used in various fields.
A New Consideration for Chief Information Security Officers
For Chief Information Security Officers (CISOs), Eset’s announcement redefines the purchasing landscape. Selecting a detection platform now involves choosing not only the software that analyzes logs but also the model that drives its underlying analysis. Eset responds with an in-house model, while competitors often utilize large AI provider models.
On a related note, Tanium, a provider of endpoint management and security solutions, announced on November 3 that it would join OpenAI’s Trusted Access for Cyber program. This initiative grants security vendors access to OpenAI’s advanced cyber capabilities for threat detection and vulnerability research. CrowdStrike has been part of this collaboration since April. In this scenario, the model provider is a third party, and the security vendor’s contract with them falls within the client’s supply chain.
Clients of Eset do not purchase NeoCortex separately, as it operates within LiveCortex. They should review their contracts to ascertain how their telemetry is processed, what actions the model can perform without human validation, and what performance evidence the company provides. The claim of being 200 times cheaper and the incident ratio of one for every 90,000 events warrant scrutiny against observed outcomes in their own environments.



