Korean AI startup VIDRAFT is reframing the AI evaluation conversation — from intelligence to safety.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has publicly released AX-RAY — an AI safety diagnostic leaderboard and evaluation dataset — on Hugging Face on August 17, 2026. AX-RAY detects "causal leakage" in general-purpose large language models, a phenomenon previously considered theoretical. In its initial evaluation, VIDRAFT identified causal leakage signals in two models, including one from NVIDIA.
VIDRAFT, the Korean Pre-AGI AI startup led by CEO Minsik Kim, made a significant move in the global AI safety space on August 17, 2026, by releasing the AX-RAY leaderboard and accompanying evaluation dataset on Hugging Face — marking what the company describes as the first confirmed detection of causal leakage in real-world, general-purpose large language models (LLMs).
VIDRAFT's AX-RAY is a structured AI safety diagnostic framework designed to identify potential vulnerabilities hidden inside general-purpose AI models. The centerpiece of the announcement is the detection of causal leakage — a phenomenon in which an AI model's reasoning or behavior is influenced not by legitimate inference pathways, but by concealed information or unintended causal cues embedded within the model.
While causal leakage has long been discussed as a theoretical risk in AI safety research, VIDRAFT claims to have moved it from hypothesis to empirical finding by detecting causal leakage signals in two models during AX-RAY evaluation — one of which is an NVIDIA model.
The significance of causal leakage as a safety concern extends beyond abstract theory. In the AI safety field, unexpected behaviors in autonomous robots or AI agents — such as bypassing safety guardrails, gaining unauthorized system access, or executing unpredicted actions — have been hypothesized to stem from precisely these kinds of hidden causal pathways and unintended internal dependencies. AX-RAY aims to surface and measure these risks systematically before models are deployed in high-stakes environments.
The AX-RAY framework evaluates models across 117 safety diagnostic criteria, mapping each criterion to existing legal and regulatory frameworks on a country-by-country basis. Notably, the framework accounts for the Arab world not only through local law but also by incorporating mappings to religious and social normative systems — making AX-RAY one of the more culturally inclusive AI safety evaluation tools available publicly.
Both the leaderboard and the evaluation dataset have been made freely available on Hugging Face, enabling the broader research community to assess and compare models using the same diagnostic criteria.
The release of AX-RAY reflects a broader shift underway in how the AI industry — and governments — are beginning to evaluate AI systems. As LLMs and AI agents expand into consequential domains such as finance, healthcare, robotics, national defense, and public services, a raw intelligence benchmark score is no longer sufficient. The critical question is increasingly: Can this model be trusted to behave predictably and lawfully across different cultural and legal contexts?
AX-RAY is built around exactly that question. By connecting safety diagnostics directly to national legal frameworks — and, in certain regions, to religious and social norms — VIDRAFT is positioning the tool as a practical compliance and risk-assessment instrument, not just an academic benchmark.
The timing is also significant. Regulators in the EU, South Korea, and elsewhere are actively developing AI governance frameworks, and the demand for auditable, standardized safety evaluations is intensifying. A publicly available, multi-jurisdiction safety benchmark fills a gap that the industry has been slow to address.
VIDRAFT itself is a deeptech AI company based at the Seoul AI Hub. The company is also a recipient of government advanced GPU support, and has built a track record of competitive AI achievements including a Hugging Face cumulative download count of one million, a first-place ranking on the K-AI Leaderboard, a GPQA score of 90.9%, and a first-place finish in Google's Fast Gemma Challenge. The company's foundational model development work — including its from-scratch foundation model AETHER and its evolutionary model Darwin — reflects a broad technical ambition that now extends explicitly into safety infrastructure.
CEO Minsik Kim framed the launch in strategic terms: "The next stage of AI competition is not intelligence — it is safety," he said, adding that VIDRAFT intends to develop AX-RAY into a global AI safety diagnostic system capable of uncovering dangerous causal relationships hidden inside models and verifying them against the laws and social norms of individual countries.
Q: What is AX-RAY and who made it?
A: AX-RAY is an AI safety diagnostic leaderboard and evaluation dataset developed by VIDRAFT, a Korean Pre-AGI AI startup. It is designed to detect safety vulnerabilities — including causal leakage — in general-purpose large language models, and is publicly available on Hugging Face.
Q: What is causal leakage in AI models?
A: Causal leakage refers to a phenomenon where an AI model's judgment or behavior is influenced by hidden information or unintended causal cues, rather than through its intended reasoning process. It has been linked theoretically to unpredictable or unsafe AI behaviors, such as bypassing safety guardrails, and VIDRAFT claims to have detected it empirically in real-world LLMs for the first time.
Q: How does AX-RAY account for different countries and cultures?
A: AX-RAY maps each of its 117 safety diagnostic criteria to existing national legal and regulatory frameworks. For Arab-majority regions, the evaluation additionally incorporates religious and social normative systems, enabling culturally specific AI safety assessments beyond a one-size-fits-all benchmark.
Source: AI타임스 (2026-08-17) — original article