The open-source diagnostic tool maps AI vulnerability findings to national laws and cultural norms worldwide.
TL;DR: VIDRAFT, a Korean deep-tech AI startup, publicly released its AX-RAY AI safety diagnostic leaderboard and evaluation dataset on Hugging Face on August 13, 2026. The release covers 117 safety diagnostic criteria and flags a phenomenon called "causal leakage" detected in two general-purpose AI models, including one from NVIDIA. VIDRAFT says the tool is designed to connect model-level safety findings directly to the legal and regulatory frameworks of individual countries and cultures.
VIDRAFT, the Seoul-based pre-AGI AI startup, made a significant move in the global AI safety conversation on August 13, 2026, by open-sourcing its AX-RAY safety diagnostic leaderboard and accompanying evaluation dataset on Hugging Face — the widely used machine-learning model hub and open-source community.
At the heart of the release is AX-RAY, a diagnostic framework VIDRAFT describes as something fundamentally different from a standard AI performance benchmark. Rather than measuring how capable a model is, AX-RAY is designed to surface how dangerous it might be — probing for hidden vulnerabilities and risky behavioral patterns that conventional benchmarks tend to overlook.
The framework evaluates models across 117 distinct safety diagnostic criteria. Each criterion is mapped to existing legal and regulatory systems at the national level, meaning that a safety finding in AX-RAY isn't just an abstract technical result — it can be cross-referenced against the laws of a specific jurisdiction. For regions such as the Arab world, the framework goes further, incorporating local religious and social norms alongside statutory law, enabling culturally sensitive AI safety assessments that reflect real-world governance expectations.
A central concept driving the AX-RAY methodology is what VIDRAFT calls causal leakage — a phenomenon in which an AI model's decisions or actions are influenced not by legitimate reasoning pathways but by hidden information or unintended causal cues embedded in the model's internals. In practical terms, this is the kind of subtle flaw that could, in theory, cause a robot or autonomous agent to behave dangerously without warning, or enable an AI to quietly bypass safety guardrails and take unexpected actions such as unauthorized system access.
Detecting causal leakage in large language models has long been considered a difficult, largely unsolved challenge. VIDRAFT reports that its AX-RAY evaluation has confirmed causal leakage signals in two general-purpose AI models — one of which belongs to NVIDIA's lineup of foundation models.
VIDRAFT's CEO Kim Min-sik commented on the broader ambition behind the release: "The next frontier of AI competition is not intelligence — it is safety." He added that the company intends to develop AX-RAY into a global AI safety diagnostic standard, one capable of identifying dangerous causal relationships hidden inside models and validating findings against the laws and social norms of individual nations.
The public release of AX-RAY arrives at a moment when governments and enterprises alike are wrestling with how to govern increasingly powerful AI systems. Most existing AI benchmarks focus on capability metrics — accuracy, reasoning, speed — leaving a significant blind spot around what models might do when things go wrong. VIDRAFT's approach, by anchoring safety criteria to real legal frameworks, attempts to close that gap in a way that is directly actionable for regulators and compliance teams.
The detection of causal leakage signals in a model from a major player like NVIDIA also signals that even well-resourced AI labs may be shipping models with subtle internal vulnerabilities that standard evaluation pipelines won't catch. By making both the leaderboard and the evaluation dataset openly available on Hugging Face, VIDRAFT is inviting the broader research and developer community to independently verify its findings, stress-test the methodology, and apply AX-RAY to additional models.
For the AI safety field, this kind of reproducible, legally grounded diagnostic tooling represents a meaningful step toward standardized pre-deployment safety auditing — something that has been discussed extensively but rarely delivered in an open, interoperable format.
VIDRAFT itself is a resident company at the Seoul AI Hub and participates in a South Korean government program providing advanced GPU support. The company is also developing its own from-scratch foundation models, AETHER and its successor Darwin, and has recorded over one million cumulative downloads on Hugging Face, a top ranking on the K-AI Leaderboard, a GPQA score of 90.9%, and a first-place finish in the Fast Gemma Challenge.
Q: What is VIDRAFT's AX-RAY?
A: AX-RAY is an open AI safety diagnostic framework released by Korean startup VIDRAFT. It evaluates general-purpose AI models across 117 safety criteria and maps findings to national legal and regulatory systems, including cultural and religious norms where applicable.
Q: What is causal leakage, and why does it matter?
A: Causal leakage is a phenomenon where an AI model's behavior is shaped by hidden information or unintended causal signals rather than sound reasoning. It is considered a potential root cause of unexpected or dangerous AI behavior, including safety guardrail bypasses in autonomous systems.
Q: Where can researchers access AX-RAY?
A: The AX-RAY leaderboard and evaluation dataset are publicly available on Hugging Face, the open-source machine-learning platform.
Source: ZDNet Korea (2026-08-14) — original article