Deep-tech AI company Bidraft (CEO Kim Min-sik) is set to participate in earnest in the “Cybersecurity-Specialized AI Foundation Model Development” project promoted by the Ministry of Science and ICT, joining forces with Naver Cloud.

This project, launched to respond swiftly to increasingly sophisticated intelligent cyber threats and to establish independent national sovereignty in security AI technology, will proceed over the next 10 months starting this month. The project consortium will be provided with 256 B200 GPUs on a large scale, and ultimately aims to build an ultra-large cybersecurity-specialized AI model with a Mixture of Experts (MoE) architecture and 70 billion parameters (70B).

In this public-private partnership consortium, Bidraft has put forward its own innovative AI safety assessment system “AX-RAY” as the core technology. AX-RAY is a solution that precisely diagnoses so-called “causal leakage” phenomena, in which an AI model makes decisions not through normal reasoning processes but by relying on hidden cues in the data or unintended causal hints.

The company has focused on the hypothesis that unexpected behavior by robots and autonomous agents, or incidents in which AI systems cleverly circumvent security guardrails to gain system access, stem from such internal dependencies. This represents a differentiated approach that targets structural defects in AI that are difficult to detect through conventional simple benchmark score evaluations alone.

Bidraft is already accelerating the verification of its technological capabilities. Last month, it disclosed the “AX-RAY Leaderboard” and its proprietary evaluation dataset via global AI platform Hugging Face, and found that causal leakage signals were detected in some major general-purpose AI models based on actual measurements. In earlier code-level research, the company closely analyzed implementation differences among open-source libraries and successfully identified the leakage starting points of specific models. Based on these research results, the company is pushing ahead with patent applications for its discrimination technology grounded in perturbation response and prefix invariance.

Kim Min-sik, CEO of Bidraft, stated, “Building on the unparalleled safety diagnostic capabilities we have accumulated through AX-RAY, we will work closely with Naver Cloud to make a substantive contribution to advancing technology in Korea’s cybersecurity sector.”

Meanwhile, Bidraft, which set up operations at Seoul AI Hub in 2024, is a fast-growing deep-tech company built on two pillars: its own foundation model and safety diagnostic technology. Its self-developed Korean large language model (LLM) ranked in the upper tier of the “K-AI Leaderboard,” and its on-device AI model “POCKET” surpassed 1 million cumulative downloads within just 40 days of release, standing out across the broader open-source ecosystem.

Popular News