According to AI TIMES, Wonderpulp announced on September 18 that it has signed an approximately 3-year global partnership with FC Bayern Munich and will fully introduce enterprise AI agents into club operations, starting with fan support on game days. The first application will be fan service that handles digital ticket and account-related inquiries. The key to the partnership is not simply attaching separate tools, but connecting the club's existing systems with AI so it can be used within actual workflows. On site, the plan is to automate repetitive inquiries to reduce waiting times and confusion, while the club aims to improve service efficiency. The scope of adoption is expected to extend beyond fan support to back-office and broader business operations. The article emphasizes that sports clubs are shifting from using AI as an analytical aid to delegating the operations themselves to AI.

According to AI TIMES, CyberLink unveiled "PowerDirector 2027" and "PhotoDirector 2027" on Sept. 18, bringing AI agents into its video and photo editing software in earnest. The new lineup also expands more than 10 generative AI features, shifting the workflow so users can describe the desired outcome in natural language and then carry out multiple steps in sequence, from editing and retouching to marketing content creation. The key change is that the design no longer simply adds AI to individual functions, but instead hands over the creative flow itself for AI to execute. In video, users can give voice commands for tasks such as organizing source materials, editing, removing objects, applying effects and generating subtitles. In photo editing, the company strengthened conversational editing to reduce complex retouching steps, and it also embedded the same kind of agents into its marketing platform to broaden their use cases. It also added an on-device AI computing option so the creative software can switch between cloud processing and PC-side processing depending on the work environment.

According to AI TIMES, Xenon unveiled an open-source action-oriented AI model, "Hunmin VLM 397B," on September 18 that directly operates by looking at and manipulating computer screens. The key feature of the model is that it was post-trained on 8 NVIDIA B200 GPUs to add computer-use capabilities to a giant open model while significantly reducing the decline in Korean-language performance. It also stood out in its ability to identify targets such as on-screen buttons and input fields, and it is aimed at carrying out tasks across actual operating systems and applications. The company said it focused on translating user intent into action rather than simply generating answers. It also released the model, an FP8 version, and its evaluation methodology together so they can be used for performance comparisons and follow-up research. Attention is now on whether a development approach that efficiently adds the necessary capabilities on top of an open model will become one pillar of expanding practical AI for work.

According to AI TIMES, Dinotisia unveiled a server configuration equipped with 4 VDPU cards in the United States, putting a dedicated processor strategy for search in generative AI infrastructure front and center. At an event held from September 15 to 17, the company said the server delivered up to 5.77 times higher vector search throughput than a CPU-only server. Dinotisia said it was extending the chip and accelerator cards it had previously unveiled into a real server environment, highlighting a division of labor in which GPUs handle model execution and VDPUs handle external information retrieval. In particular, it emphasized that it raised throughput without degrading search quality, targeting use cases such as retrieval-augmented generation and AI agents. In South Korea, the company is pursuing productization as enterprise AI data infrastructure bundled with its own database, while overseas it is expanding evaluation and verification scope among related industries. The company also said the unveiling would be a starting point for making the applicability of server-type VDPU more concrete based on actual customer workloads.

According to AI TIMES, IBK Investment & Securities said on September 17 that Lotte Innovate's data center DBO business has entered a growth phase and could account for up to 30% of the company's total sales in 2028. The firm said the company is expected to benefit as rising investment in domestic AI data centers drives demand for one-stop services covering design, construction, and operations. It also said Lotte Innovate has built competitiveness in a structure where it handles the full process for customer-owned facilities, drawing on its long operational experience. Recently, the company has also been boosting the business's presence by securing a series of outsourced operations and DBO contracts. In addition, the expanding participation of investors in data center development was cited as a favorable environment for business expansion, while expectations for improved performance at subsidiaries have also raised hopes that the company's overall earnings strength will improve as well.

According to AI TIMES, Hancom stepped up its push into the U.S. market on September 18 by unveiling the first TV ad for its AI workforce platform, Nomadian. The ad features five AI agents with different work roles appearing like team members, while a solo founder runs a company with them. Through this, Hancom presented AI not as a simple tool but as a teammate that works alongside users, and showed a way to build a team by choosing and deploying agents and, if needed, creating them directly. In particular, it highlighted as a differentiating factor that the agents grow to resemble the user's own way of working as conversations and feedback accumulate. Starting with domestic broadcasts, the ad will also be used in marketing aimed at U.S. users, and the service is set to move into U.S. beta release and commercialization. The move is seen as a strategy to build a stronger presence in the solo founder market by tying the brand message to the experience of working with AI itself rather than to simple feature competition.

According to AI TIMES, UNIST was selected on September 18 for a project to develop core security technology that connects judgment and execution in agentic AI, and will receive about KRW 11 billion in funding. Together with Sungkyunkwan University, it will conduct research through December 2031, with the core goal of building an integrated defense system that prevents external attacks from leading to information leaks or unauthorized tasks. The research will move beyond an approach that protects only the model and focus on designing both the interfaces through which inputs enter and exit and the actual execution environment. The plan is to link safeguards so that even if one stage is breached, attacks can be blocked at the next stage, responding to threats that pass through multiple stages. It also plans to verify the technology in the actual services of partner companies and assess whether it can be applied on-site. The project also carries significance in lowering the barrier to adoption in areas where deployment had lagged because of security concerns and in broadening the research workforce that understands both AI and security.

According to AI TIMES, the Ministry of SMEs and Startups selected 40 winners after the second round of presentation evaluations for the "Everyone's Startup: Social Innovation Social Venture League," and the final 10 will be decided at a public presentation evaluation during "Venture Week 2026" in December. The selections, announced on September 18, were based on judging focused on how the entrants approach social problem solving. The program aims to identify founders and nurture them into social ventures in areas such as care gaps, jobs and social participation, energy efficiency, productivity improvements using AI, and resource circulation. The successful candidates will further refine their solutions and test their feasibility in the field with the help of private startup planners and investment firms. Follow-up support will include investment, marketing, sales channel development, technology commercialization, expert consulting, and access to public data. The ministry said it will continue supporting the effort so that startups addressing social problems can serve as a stage where innovation and growth potential are proven together.

According to AI TIMES, MIT researchers unveiled an AI technology called XvRA on September 16 that adapts to patient-specific CT and MRI scans in about 5 minutes and then precisely aligns intraoperative X-rays with 3D medical images. The technology links a patient's 3D anatomical information to flat X-rays, helping clinicians identify tool positions more accurately in minimally invasive surgery, and its registration accuracy was reported at the submillimeter level. Previously, medical staff had to manually set reference points or align positions, which placed a heavy burden on them, but this approach aims to reduce that process with patient-specific AI. The key is not to use one general-purpose model for all patients, but to quickly fine-tune separate models to each patient's unique anatomy. The researchers said evaluations using real hospital data showed that it worked stably across different patients, body parts, and medical procedures. Going forward, the challenges are to expand its use to emergency procedures and surgical robot navigation, and to extend it to moving organs and tissues.

According to AI TIMES, a report unveiled on September 17 by EverPure and Omdia found that more than one-third of enterprise data remains unused dark data, identifying it as a key risk that undermines AI readiness. The survey showed that 97% of organizations are struggling to move AI pilots into live operations. The report argued that simply accumulating more data makes it hard to deliver AI results, and that duplicated, outdated, or unnecessary information mixed in with valuable data adds to costs as well as security and regulatory burdens. It also said companies recognize the risks of dark data but still lack sufficient visibility and control across their data environments. It suggested that businesses need a way to assess data value and risk together and decide which information to use immediately, which to manage strictly, and which to reduce or eliminate. In the end, the article raises the issue that AI competitiveness depends less on the model itself than on whether a company has an operating system that can explain where its data is stored, who uses it, and how trustworthy it is.

According to AI TIMES, a Gyeongnam delegation visited the Yangtze AI Robot Valley in Shanghai, China, on Sept. 17 and conducted an on-site review for the creation of a "manufacturing AI demonstration and diffusion hub." The delegation visited AGIBOT, a humanoid company that has grown rapidly in just 3 years since its founding, and examined a production system that operates robots and AI models together, as well as examples of their application on manufacturing sites. The visit focused on how the ecosystem of a world-class robot industry cluster takes shape and how technology is introduced into actual processes. Gyeongnam Province said it plans to flesh out practical measures to spread AI transformation across manufacturing in the province based on the visit. It also reiterated its vision of creating an evidence-based hub by combining on-site data with its existing manufacturing base in aerospace, defense, and nuclear power. On the same day, the delegation also carried out a tourism cooperation schedule, broadening ties with China in both industrial and consumer areas.

According to AI TIMES, Nokia and MS said on September 17 that they will jointly build an integrated data foundation that allows telecom operators to use network data within minutes, accelerating the shift to AI agent-centric autonomous networks. The plan is to significantly reduce the data integration and preparation process, which previously took weeks, and expand AI-based automation to root-cause analysis of outages, predictive maintenance, and closed-loop operations. At the center of the partnership is the effort to combine telecom-specific data products with an integrated data and analytics platform to unify fragmented data environments. Based on this, the system is designed to increase network autonomy even in multi-vendor environments and in on-premises, cloud, and hybrid environments. They also said AI would function as a copilot to assist operations rather than replace engineers, with the scope of automation expanded step by step under human supervision. Initial use cases include voice service quality assurance, location-based experience analysis, and fault management, and the key question is whether real-world deployment will expand further.