[Edaily Reporter Kim Hyun-ah ] Naver (NAVER(035420)) is expanding the scope of its AI safety management from AI models to its entire range of actual services. Since the risks associated with the same AI model vary depending on which service—such as search, shopping, or weather—it is applied to, the company has formalized a system to identify and manage risks tailored to the context of each specific service.

With the implementation of the Framework Act on Artificial Intelligence in January 2026 and the growing prevalence of agentic AI—where multiple AI models are combined and operate autonomously—Naver has moved beyond declarative ethical principles to specify concrete safety standards for actual services.

NAVER announced on the 16th that it has published the “NAVER AI Safety Progress Report 2026,” an official report detailing real-world service implementation cases of its proprietary AI safety framework, “NAVER ASF 2.0,” and its execution system, “CHEC 2.0.”

Following the “NAVER AI Ethics Guidelines” in 2021 and “ASF Beta” for frontier AI models in 2024, the company has now expanded the scope of AI safety management from models to users and services as a whole.

AI Risks Categorized into 110 Subcategories

In this report, NAVER presented 10 major safety management frameworks that operate across the entire lifecycle—from model development to service deployment and post-deployment monitoring.

① N-ARTI: Classifies legal, social, and industrial AI risks into 110 specific categories

② AI Impact Assessment: Differentiated management of “high-impact AI” special domains and general domains under the Framework Act on Artificial Intelligence

③ Crisis Response: When users make crisis-related inquiries—such as those regarding self-harm—the system connects them to professional counseling agencies rather than simply rejecting the request

④ Risk Updates: Dynamic updates that incorporate new legal and social issues into the classification system within three days

⑤ N-ASET: Inputting 10,000 hostile queries per round, both before and after launch

⑥ Prevention of Excessive Rejections: Treating the phenomenon of unconditionally blocking even normal questions as a service defect

⑦ Disclaimer Management: Quantitatively monitor overreactions, such as the excessive use of unnecessary disclaimers in general queries

⑧ CHEC 2.0: Operation of a five-step process: Identification → Consultation → Mitigation → Measurement → History Management

⑨ Three-tier accountability system: Governance structure extending from operational teams → AI Safety Center (AISC) → Board of Directors

⑩ User Communication: Display the “On-Service AI” logo and provide information on the limitations of AI responses and data sources

The starting point here is to specifically categorize the risks that AI can pose.

Based on case studies in the legal, social, and industrial sectors, Naver has developed “N-ARTI (NAVER AI Risk Taxonomy & Identification),” which classifies potential risks into 110 specific categories.

These include inquiries regarding crisis situations such as self-harm; unqualified advice in specialized fields such as medicine, finance, and law; incitement to illegal acts; violations of laws and regulations related to food labeling and advertising; discriminatory or hateful speech; harmful content targeting minors; personal information breaches; and emotional over-engagement with AI.

The social impact of the service is also assessed in advance. Using the “AI Impact Assessment Matrix,” the company distinguishes between special domains—where “high-impact AI” as defined by the Framework Act on Artificial Intelligence may be utilized—and general domains, and adjusts the rigor of safety measures according to the risk level.

The classification system is also updated promptly as new risks emerge. Naver explained that when an issue involving violations of laws and regulations regarding food labeling and advertising was identified, it added a new risk category within three days.

Testing with 10,000 attacks at a time… Checking for “over-rejection” as well

Actual safety verification is handled by “N-ASET (NAVER AI Safety Evaluation Toolkit),” the company’s proprietary evaluation tool.

Naver assesses AI safety through pre-launch evaluations and regular post-launch evaluations. Through red teaming—which involves feeding the AI 10,000 adversarial queries per evaluation session—the company verifies whether the AI produces unexpected dangerous responses or exhibits vulnerabilities.

Safety evaluations are not limited to dangerous responses.

“Excessive refusal”—where the AI fails to answer even normal questions—is also viewed as an issue that degrades service quality and is managed accordingly. The company also checks for “excessive responses,” such as the repeated inclusion of unnecessary disclaimers in answers to general questions.

The goal is to avoid compromising the AI’s usefulness by focusing solely on risk avoidance. This is why Naver emphasizes “Unfelt Safety.”

Management from Service Planning Through Post-Launch

These safety standards are applied from the planning stage

through post-launch

via “CHEC 2.0,” the company-wide execution framework.

Service teams and the AI Safety Center collaborate to identify risks, discuss mitigation strategies, measure outcomes, and manage records through a five-step process. The company has also established a three-tier management and oversight system linking operational teams, the AI Safety Center, and the Board of Directors.

A real-world example of this is the conversational search feature “AI Tab.”

Officially launched last June, AI Tab surpassed 10 million cumulative users within a month. It utilizes a “product-native LLM” based on MoE (Mixture of Experts), optimized for large-scale service environments.

The roles are divided such that a fast, small-scale model first derives a conclusion, while a large-scale model generates the body text. Furthermore, it is designed to provide answers based on information explicitly stated in Naver’s search database to reduce the generation of incorrect information, and in specialized fields such as healthcare, it directs users to expert advice.

Naver is expanding this same principle to generative AI services such as the AI Shopping Agent and Naver Weather. It is also strengthening communication with users by displaying the “On-Service AI” logo and providing information on the limitations of AI responses as well as the sources of data used, such as the Korea Meteorological Administration and public data.

Expanding Safety Research with KAIST, Seoul National University, and Others

Naver is also expanding its efforts to validate AI safety standards and technologies in collaboration with external research institutions.

A paper on “Stable-GFlowNet,” a red-team technology jointly developed by Naver AI Lab and KAIST, was selected for the “Spotlight” category at ICML 2026—an honor reserved for the top 2.2% of papers. This technology identifies vulnerabilities in AI models prior to deployment by exploring a wider variety of attack patterns than existing methods.

During the planning and drafting of ASF 2.0, Naver collaborated with Seoul National University’s Artificial Intelligence Policy Initiative (SAPI) for expert advisory support. It is also collaborating with Korea University’s Artificial Intelligence Safety Research Center and shared real-world application cases at AI safety events organized by the Ministry of Science and ICT and the Artificial Intelligence Safety Institute (AISI).

At “SAIPCON 2026,” hosted by Seoul National University’s SAPI, Naver Cloud Center Director Noh Sang-min addressed the topic “AI Data Centers and Power Grids: How Will They Coexist?” highlighting the power demand of AI infrastructure and bottlenecks in the Seoul metropolitan area’s transmission grid. He also proposed the need for institutional design measures, such as directing AI data centers toward regions with surplus power grid capacity.

In addition, the Korean edition of the “International AI Safety Report 2026” featured Naver’s multilingual, service-centric AI safety assessments and its participation in the Singapore IMDA Global Red Teaming Challenge. Naver is also working toward obtaining ISO/IEC 42001 certification, the international standard for AI management systems.

Kim Myung-ju, Director of the Artificial Intelligence Safety Research Institute, stated, “Laws merely outline the minimum standards; it is up to the companies creating the services to determine how to address specific risks within that framework.” He added, “This report documents Naver’s first practical effort to build safety at the service level—rather than at the model level—on its own.”

Won Seong-jae, Head of Naver’s AI Safety Center, said, “As AI technology becomes deeply embedded in our daily lives, AI safety has moved beyond theoretical discussion and entered a phase where it must be addressed in actual service environments.” He added, “We will continuously enhance ASF 2.0 to prioritize human-centered values throughout the entire process of AI development and use, and to prevent and mitigate risks that may arise in our services.”

Yoo Bong-seok, Naver’s CRO, remarked, “AI safety is not an obstacle to innovation but rather the foundation of innovation that helps more people trust and use AI,” adding, “We will transparently share the process of finding better solutions together with society.”