Key Takeaways
- • KT and Korea University jointly released KSAFE-MM on June 16, 2026, a multimodal AI safety benchmark comprising 14,135 evaluation samples designed specifically for Korean language and cultural context.
- • KSAFE-MM consists of two components: KSAFE-MM-G (converting global safety risks into Korean cultural context) and KSAFE-MM-C (addressing Korea-specific issues including jeonse rental deposit fraud and the Dokdo territorial dispute).
- • The benchmark was validated across 12 global multimodal large language models, including Gemma and HyperCLOVA X, establishing a standardized safety evaluation framework for AI systems operating in the Korean market.
- • KT developed an automated four-step pipeline for benchmark creation (sensitive topic collection, query generation, synthetic image generation, and jailbreak query generation) that reduces manual review costs and enables rapid localization to other cultural contexts.
- • The research and benchmark were published on arXiv and Hugging Face; a pilot expansion (JSAFE-MM-C) demonstrates the pipeline's scalability to other markets, particularly Japan.
Summary
KT unveiled KSAFE-MM on June 16, 2026, a culturally localized safety benchmark for multimodal large language models developed in partnership with Korea University. The benchmark addresses a critical gap in AI safety evaluation by incorporating Korean language requirements and society-specific risks that global benchmarks typically overlook. This release reflects South Korea's broader strategic positioning in AI governance, particularly following the country's implementation of the world's first comprehensive AI Basic Act.
KSAFE-MM comprises 14,135 evaluation samples organized into two distinct categories. KSAFE-MM-G adapts universal safety concerns—such as bias, toxicity, and misinformation—to the Korean cultural and linguistic environment. KSAFE-MM-C addresses localized risks unique to South Korean society, including jeonse fraud (rental deposit schemes) and the Dokdo dispute (territorial disagreement with Japan), which require cultural and geopolitical context to evaluate properly. This dual-component approach ensures that AI systems deployed in Korea can be assessed against both international safety standards and locally relevant risk factors.
The validation process tested the benchmark against 12 prominent multimodal large language models, including Gemma and HyperCLOVA X, confirming its practical utility across diverse AI architectures. KT's development of an automated pipeline represents a methodological advancement, moving away from labor-intensive manual review processes. The four-step pipeline—covering sensitive topic identification, template-based query generation, synthetic image generation, and jailbreak query generation—significantly reduces the time and expertise required to create culturally specific benchmarks. A pilot expansion to Japanese (JSAFE-MM-C) demonstrates the framework's potential for rapid replication across other markets and cultural contexts without requiring domain experts from each region.
The publication of KSAFE-MM on both arXiv and Hugging Face reflects KT's commitment to advancing industry standards through open-source distribution. This development positions South Korea as a leader in specialized AI safety infrastructure and aligns with the country's regulatory framework, which increasingly emphasizes trust and controllability as competitive differentiators in the autonomous AI agent market.
KSAFE-MM comprises 14,135 evaluation samples organized into two distinct categories. KSAFE-MM-G adapts universal safety concerns—such as bias, toxicity, and misinformation—to the Korean cultural and linguistic environment. KSAFE-MM-C addresses localized risks unique to South Korean society, including jeonse fraud (rental deposit schemes) and the Dokdo dispute (territorial disagreement with Japan), which require cultural and geopolitical context to evaluate properly. This dual-component approach ensures that AI systems deployed in Korea can be assessed against both international safety standards and locally relevant risk factors.
The validation process tested the benchmark against 12 prominent multimodal large language models, including Gemma and HyperCLOVA X, confirming its practical utility across diverse AI architectures. KT's development of an automated pipeline represents a methodological advancement, moving away from labor-intensive manual review processes. The four-step pipeline—covering sensitive topic identification, template-based query generation, synthetic image generation, and jailbreak query generation—significantly reduces the time and expertise required to create culturally specific benchmarks. A pilot expansion to Japanese (JSAFE-MM-C) demonstrates the framework's potential for rapid replication across other markets and cultural contexts without requiring domain experts from each region.
The publication of KSAFE-MM on both arXiv and Hugging Face reflects KT's commitment to advancing industry standards through open-source distribution. This development positions South Korea as a leader in specialized AI safety infrastructure and aligns with the country's regulatory framework, which increasingly emphasizes trust and controllability as competitive differentiators in the autonomous AI agent market.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.