The Korean AI startup goes public with a cross-cultural model evaluation framework — and flags causal leakage signals in two widely used LLMs.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has publicly released its AI safety diagnostic framework AX-RAY as an open leaderboard and evaluation dataset on Hugging Face. The system evaluates AI models across 117 diagnostic items and detected causal leakage signals in two general-purpose AI models, including one developed by NVIDIA. VIDRAFT plans to expand AX-RAY coverage to LLMs and AI agents deployed in finance, healthcare, robotics, and public services.
VIDRAFT, the Korean Pre-AGI AI startup led by CEO Kim Min-sik, publicly released its proprietary AI safety diagnostic system, AX-RAY, on August 18, 2026 — making its evaluation leaderboard and accompanying dataset freely available on Hugging Face. The release marks a significant step in the company's push to establish a globally applicable standard for AI safety assessment, one that goes beyond raw model performance to scrutinize how AI systems behave when exposed to hidden causal cues.
The AX-RAY framework evaluates AI models across 117 diagnostic items designed to surface potential risk behaviors and safety vulnerabilities that may not be visible through conventional benchmarks. What distinguishes AX-RAY from existing evaluation tools is its explicit linkage of safety findings to the legal and regulatory frameworks of specific countries and regions.
For example, when evaluating models intended for use in Arabic-speaking countries, AX-RAY incorporates not only local statutory law but also religious and social norms — reflecting the reality that AI safety standards differ meaningfully across cultures and jurisdictions. This culturally-grounded approach positions AX-RAY as a tool for global AI governance rather than a narrowly technical benchmark.
The centerpiece finding of the initial AX-RAY evaluation is the detection of causal leakage signals in two publicly available general-purpose AI models. One of those models was developed by NVIDIA. According to VIDRAFT, causal leakage refers to a phenomenon in which an AI system's judgments are influenced not by a legitimate reasoning pathway but by hidden information or unintended causal cues embedded in the model's training or architecture. While the concept has circulated as a theoretical risk in AI safety discourse, VIDRAFT says that reliably detecting such signals in production-grade large language models has until now remained elusive.
Importantly, VIDRAFT emphasizes that the detected signals indicate potential vulnerabilities within specific evaluation items — not evidence that these models have already caused harmful outputs or compromised any system. The distinction matters: AX-RAY is designed to flag latent risk before it becomes realized harm.
The release arrives at a moment when regulators in the EU, the United States, and across Asia are scrambling to define what responsible AI deployment actually looks like in high-stakes sectors. Most existing leaderboards rank models by capability — reasoning accuracy, code generation, or language fluency. AX-RAY inverts that priority, treating safety as the primary axis of evaluation.
By hosting the leaderboard openly on Hugging Face and making the evaluation dataset publicly accessible, VIDRAFT is inviting the broader AI research community to scrutinize and stress-test its methodology. This openness also allows organizations deploying LLMs in regulated industries — banking, clinical care, public administration, autonomous systems — to use AX-RAY as a pre-deployment screening tool.
CEO Kim Min-sik framed the ambition in direct terms: "The next stage of AI competition is not intelligence — it is safety. We will develop AX-RAY into a global AI safety diagnostic system that uncovers dangerous causal relationships hidden inside models and validates them against the legal frameworks and social norms of each country."
VIDRAFT intends to keep expanding AX-RAY's scope, adding more models to the leaderboard and broadening the diagnostic dataset. The longer-term goal is a comprehensive safety vetting system for any LLM or AI agent operating in finance, medicine, robotics, or government services — essentially, any domain where a miscalibrated causal inference could carry real-world consequences.
Q: What is AX-RAY and who developed it?
A: AX-RAY is an AI safety diagnostic framework created by VIDRAFT, a Korean Pre-AGI AI startup. It evaluates AI models across 117 items to detect potential risk behaviors and safety vulnerabilities, and connects findings to relevant national laws and cultural norms.
Q: What is causal leakage, and why does its detection matter?
A: Causal leakage occurs when an AI model's outputs are influenced by unintended hidden cues rather than sound reasoning. Detecting it early matters because it represents a latent vulnerability that could lead to unreliable or unsafe behavior in high-stakes deployments — even if no harm has yet occurred.
Q: Where can researchers and developers access the AX-RAY leaderboard and dataset?
A: Both the AX-RAY leaderboard and its evaluation dataset are publicly available on Hugging Face, allowing the global AI community to review model safety rankings and use the dataset for independent research or pre-deployment testing.
Source: IT조선 (2026-08-18) — original article