SK Telecom and Upstage Team Up to Train AI for Cyberattack Response
Eleven organizations, including six security firms and three universities, have joined the consortium. The plan: train A.X K and Solar on real-world security data, verify that they are safe, and deploy them in security products.
SK Telecom and Upstage are teaming up to train AI that assists with security work: finding and fixing weaknesses in software, and responding to hacking incidents. On October 8, SK Telecom announced the launch of its “Sovereign AI Cybersecurity Consortium” and the signing of a technology cooperation agreement. The signing ceremony was held the previous afternoon at SKT Tower.
Eleven organizations are taking part in total: the two companies, six security firms, and three universities. SK Shieldus, SECUI, AhnLab, Raonsecure, Genians, and PIOLINK are contributing field data and technology. Korea University, Soongsil University, and Pusan National University will evaluate the models’ performance and safety.
What the models will learn from security field data
The collaboration centers on SK Telecom’s A.X K series and Upstage’s Solar. Both models advanced to stage 3 of the government’s “Sovereign AI Foundation Model Project.” The consortium plans to extend the two models so they can also handle cybersecurity work. After training on data from the security firms, the models will go through university validation, with their effectiveness then confirmed in the participating companies’ products and services.
Training is split into two phases. The first builds foundational knowledge; the second raises practical, on-the-job capability. For the next model, A.X K3, the team will add security knowledge plus data aimed at operating the “AI for Everyone” service safely. The next version of Solar will be able to analyze large volumes of code and logs spanning long periods in a single pass. On top of that, Upstage will concentrate its training on data for discovering vulnerabilities. The models will then learn to find and fix software weaknesses and to handle incident response in a security operations center.
The sheer volume of accumulated field data is another foundation for this partnership. According to SK Telecom, SK Shieldus analyzes 17 billion threat signals per day in real time through its Secudium platform. AhnLab holds more than 2.5 petabytes (PB) of security data. The consortium expects that feeding the threat intelligence continuously piling up in the field into training will strengthen the models’ ability to respond even to novel attacks.
Pitting the models against each other to find weak spots
For safety validation, the two models will attack and evaluate each other. Examples of vulnerabilities uncovered in that process will be fed back into the training of both models. Earlier, while developing A.X K2, SK Telecom used automated attack testing to generate roughly 15,000 attack-response data points for training. This time, it is extending that approach into mutual evaluation between the two models. Upstage will combine its own 210,000 challenging and safety-related training examples with a training framework that teaches the model to refuse dangerous requests.
The universities will design attack tests independently of the developing companies and assess the results. Korea University will examine the models’ ability to analyze executable files and their resistance to attacks. Soongsil University will establish security performance evaluation criteria and check how well the models hold up against attacks that bypass AI safety guardrails. Pusan National University will validate the performance and safety of the next-generation models and smaller-scale models.
How it will be used in actual security products
The target for deployment is the participating companies’ security products and services. The firms plan to use the two models for tasks such as analyzing vulnerabilities for which no patch has yet been released and monitoring threats in real time. Performance and safety will be verified in live operating environments, and the analysis and response experience gained there will be folded back into training.
This partnership is an attempt to connect hands-on experience from Korea’s security front lines to AI training and product deployment. The key question is how much AI can actually help security staff with vulnerability analysis and incident response. The practical payoff will likely hinge on the task performance demonstrated in the joint pilots ahead, and on the independent safety evaluations from the universities.