Eset Launches NeoCortex: A New AI-Powered Cybersecurity Model
Slovak cybersecurity firm Eset has unveiled NeoCortex, a dedicated artificial intelligence (AI) model trained exclusively on its proprietary security data and deployed on its own servers. Rather than opting for a large, general-purpose AI model, Eset has developed a smaller, specialized model designed to predict the next steps in a cyberattack. This approach ensures that client telemetry remains within Eset’s control, with critical decisions still resting in the hands of human analysts.
Eset introduced NeoCortex on October 30 and subsequently announced it in France on November 3. This foundational model serves as the backbone for LiveCortex, the engine that correlates security signals within Eset’s enterprise protection platform, Eset PROTECT. Eset describes NeoCortex as a “model of the world,” a framework that learns how an information system behaves and how its events unfold.
This launch fulfills a promise made in May at the Eset World conference in Berlin, where the company pledged a €40 million investment in AI by 2026, including the development of foundational models tailored to cybersecurity data. LiveCortex was presented as a nascent system at that time. On October 1, Eset further enhanced its platform with controls for deployed AI agents.
A Specialized Model Built and Hosted In-House
NeoCortex is trained on past incidents, helping reconstruct missing context for incidents, summarizing complex attacks, and identifying abnormal events within sequences. According to Eset, LiveCortex processes billions of telemetry events daily, with only 0.00111%—approximately one in 90,000—being escalated to clients as incidents. This filtering aims to alleviate alert fatigue for security teams overwhelmed by too many signals.
Eset emphasizes the value of specialization for analytical work. While a large general-purpose language model may provide plausible responses across various topics, a security investigation necessitates consistent tracking of entities, timestamps, indicators of compromise, and relationships between processes throughout an incident. NeoCortex operates strictly within this specialized vocabulary, forgoing broader conversational capabilities.
The choice of a specialized model also reflects cost considerations. Eset’s internal research indicates that a model like NeoCortex can be trained at a cost up to 200 times lower than that of a large general-purpose reasoning model. Eset attributes this efficiency to training focused on cybersecurity language, reduced computational demands, and its proprietary hardware. However, this figure remains based solely on Eset’s claims, as no published protocol verifies it.
A second key argument is client data control. Developed and operated on Eset’s internal infrastructure, NeoCortex keeps sensitive client telemetry and model operations under the publisher’s control. This minimizes reliance on third-party AI providers and reduces exposure to possible business and geopolitical shifts affecting those suppliers. Juraj Malcho, Eset’s CTO, asserts that this cybersecurity intelligence “remains under our control” and enhances human expertise rather than replacing it.
Human Analysts: The Sole Decision-Makers for Critical Responses
Eset’s most ambitious promise lies in predictive capabilities, as NeoCortex aims to forecast an attacker’s next likely move, allowing defenders to intervene proactively. This strategy, termed “J-1 defense,” focuses on anticipation rather than guaranteeing a 24-hour response time. Eset cited the week leading up to October 25, dubbed “Black October”, when critical vulnerabilities in PHP, Linux, and GitLab necessitated swift, prioritized responses.
However, this predictive capability alone does not mandate actions; Eset clarifies that NeoCortex is not the sole decision-making entity, operating instead alongside specialized models, deterministic controls, correlation rules, and human expert supervision. The model operates in restricted sandboxes, utilizing read-only tools, implementation limits, continuous monitoring, and an emergency shut-off switch, ensuring that critical decisions and responses remain the responsibility of analysts.
This approach echoes CrowdStrike’s recent announcement on November 3, introducing a shut-off switch for AI agents that deviate from their intended missions. Both companies aim to incorporate robust safeguards into the governance of their AI tools, whether for defensive mechanisms or deployed assistants within organizations.
A New Consideration for CISOs
For Chief Information Security Officers (CISOs), this announcement shifts the purchasing landscape. Opting for a detection platform now also involves selecting the model that analyzes log data and the environment in which it operates. Eset has chosen a proprietary model hosted in-house, while other publishers rely on AI models from major providers.
Tanium, a terminal management and security vendor taking the opposite approach, announced on November 3 its membership in OpenAI’s Trusted Access for Cyber program. This initiative grants security publishers access to advanced cyber capabilities from OpenAI for threat and vulnerability research. CrowdStrike has been part of this framework since April. In this scenario, the model provider is a third party, and the security publisher’s contract with them becomes part of the client’s outsourcing chain.
Clients of Eset do not purchase NeoCortex separately, as the model operates within LiveCortex. They should confirm in their contracts where their telemetry is processed, what actions the model can initiate without human validation, and what evidence Eset provides regarding its performance. The claimed cost efficiency of 200 times less and the ratio of one incident per 90,000 events warrant scrutiny against observed outcomes in their own environments.



