According to AI TIMES, Everpure said on September 17 that the success or failure of AI adoption depends less on the model than on where data is stored and who controls it, stressing that data sovereignty must be reflected from the design stage, not after deployment. The presentation pointed out that 97% of organizations struggle to move AI projects into the operational stage, with dispersed and unprepared data cited as a major underlying cause. In environments where corporate data is scattered across multiple systems, departments, and countries, the same information can exist in different versions, and standards for use and protection can diverge. The report highlighted that in heavily regulated sectors such as finance, the public sector, and healthcare, companies must consider not only data location but also ownership, access rights, and jurisdiction. It also explained that having a data center in Korea or applying encryption does not end the sovereignty issue, and that contracts, legal structures, and geopolitical variables must be examined together. Ultimately, the message was that as AI spreads, companies need to design from the outset a system for finding and classifying the data they need, along with an architecture that is not locked into a specific platform and can be moved elsewhere.

According to AI TIMES, the British Army has completed a swarm-flight experiment with 8 unmanned aerial vehicles in a real military environment, confirming the battlefield potential of autonomy technology that operates multiple drones as a single cooperative group. The Defense Science and Technology Laboratory under the U.K. Ministry of Defence said on September 16 that the test was part of a software-defined swarm concept demonstration project and that it assessed swarm control performance over 8 weeks of operations. The core of this proof of concept is to move away from the method of separately controlling individual aircraft and instead integrate and operate multiple unmanned systems by mission. Different swarm systems built by British defense companies took part in the test, and the test infrastructure supporting them has been handed over to an Army operating unit and is being used for hands-on flight verification. In particular, the structure was designed with an open architecture so it can be extended beyond unmanned aerial vehicles to ground and surface unmanned systems. The British Army plans to conduct additional tests and field validation to determine which operations this technology is suited for, and then reflect that in force development and unmanned systems introduction plans.

According to AI TIMES, Inha University held the inauguration ceremony for its 17th president, Choi Ki-young, on September 16, officially launching its new presidential leadership structure. President Choi began his term on September 1, and through the event, the university signaled its leadership transition and operational direction for the next 4 years both inside and outside the campus. The ceremony included the presentation of the appointment letter, the handover of the university flag, the inauguration address, and the screening of congratulatory videos. The new president was introduced as someone who has built experience in university administration through major internal posts, along with a research career in aerospace engineering. University members and outside attendees celebrated the new start together and expressed expectations for future change. President Choi presented a direction centered on student growth, saying he would lead innovation in education and research while adding flexibility and agility on top of the university's strengths.

According to AI TIMES, MIT researchers said on September 14 that they have developed an algorithm called "HardFlow" that ensures generative AI must follow safety and physical rules in its final output. Instead of imposing constraints at every intermediate step of generation, the method is designed to apply strict conditions only to the final output, allowing the system to search for better solutions while maintaining safety. The researchers said they tested it on robot manipulation, maze navigation, and text-based image editing, and that the results consistently improved in quality compared with existing methods. In fields where safety is critical, a result that does not violate the rules takes priority over a merely plausible answer, and HardFlow is seen as an approach that addresses that boundary while preserving freedom in the generation process. The researchers also highlighted that it can be attached to already trained generative AI at the deployment stage. Because it can be applied without retraining, it is drawing attention for whether it can push the practical commercialization of generative AI one step further as it deals with real-world constraints.

According to AI TIMES, ABB and BCG on September 15 proposed a hybrid power grid that uses both AC and DC, along with an 800VDC distribution architecture, as a practical alternative for easing power bottlenecks in AI data centers. The report said that data centers need to redesign their internal power paths to reduce conversion losses and infrastructure burdens that have grown as AI servers have become more power-hungry. The key is not a shift that completely pushes out AC, but a model in which AC handles transmission and external distribution, while DC takes a larger share of high-power loads inside facilities. It said this would reduce unnecessary conversion steps and create room to operate computing resources more efficiently within limited grid connection capacity. The report also noted that this trend could spread beyond data centers to manufacturing facilities and commercial buildings. However, it said the main obstacles to adoption are less about the technology itself than about fragmented standards and a shortage of personnel able to handle them.

According to AI TIMES, Google proposed, through U.S. Patent No. 12731392, a method for retraining computer vision models so they are not trained only to match ground-truth data but are instead rewarded for actual work performance. The patent, introduced on September 14, focuses on fine-tuning after pretraining with reinforcement learning so that the model directly targets evaluation criteria that matter in the field. The key idea is that accurately mimicking labels in training data and performing well in real-world environments can be different things. Accordingly, it opens a path to linking training with criteria that better reflect on-the-ground performance across tasks such as object detection, image segmentation, and colorization. In particular, it notes that a reward function does not have to be differentiable, arguing that human judgments or composite performance assessments can also be pulled into training objectives. The article interprets this as a sign that the alignment trend in generative AI is spreading to computer vision and physical AI, and says that in the future, competitiveness may depend less on the size of the model itself than on how well goals are designed as rewards.

According to AI TIMES, Satya Nadella said on September 14 that superintelligence is not worth pursuing if it escapes human control or does not benefit humanity, warning as well against the trend of AI leadership being concentrated in a few companies. He stressed that this did not mean slowing development itself, but rather building safety and alignment into the design stage, and that companies should control their own knowledge and data within an ecosystem where closed and open-source models coexist. In particular, he said companies should not stop at simply adopting external models, but should build iterative structures for their own training and evaluation. His logic was that this would allow business knowledge and tacit know-how to be continuously reflected in models without becoming dependent on a specific provider. He also said AI safety should be treated as a design goal from the start, not as a post hoc safeguard, and that systems are needed to continuously check whether operations align with human intent. Microsoft also said it would prepare for a public consultation on its code of conduct in line with that direction, sending the message that who controls AI training direction and evaluation standards has become as important as performance competition.

According to AI TIMES, OpenAI’s Sam Altman publicly said on September 14 that the goal is not to stop frontier AI development, but that development speed can be adjusted for safety verification and monitoring. He said that even under intense competitive pressure, improvements in AI capabilities should not outrun safety safeguards, and argued that verification should extend beyond post-development checks to the training process itself. This lines up with earlier industry talks about pacing and is spreading in a direction that goes beyond outside standard-setting bodies to independent verification inside development itself. The key point is that companies should first make common safety standards and inspection systems visible rather than wait for regulation. The article highlights a shift in frontier AI competition, where the benchmark is becoming not only performance and launch speed but also the ability to prove safety. It also leaves open the next issue of what standards and accountability structure will actually be created, beyond the force of the declaration itself.

According to AI TIMES, Anthropic disclosed real-world abuse cases in a report released on September 10 showing that AI is being used not just as a coding assistant for cyberattacks, but as a coordinator spanning reconnaissance, intrusion, exfiltration, and analysis. Based on cases identified from December 2025 to August 2026, the report highlighted threats across 7 areas: cyberattacks, influence operations, surveillance, fraud, weapons development, biology, and illegal distillation. In particular, it said an automated attack loop is becoming a reality, with multiple agents handling the entire attack process in parallel and then modifying and redeploying detected malicious tools. It added that abuse has expanded to fake news production and distribution, the design of surveillance systems at a national scale, fraud combined with human workers, and support for weapons software development. It also said that in more and more fields, safety measures that only block model responses are making it difficult to distinguish legitimate research from dangerous use. The report concluded that the focus of AI security is shifting from the model itself to ecosystem management that looks at accounts, access rights, agents, external tools, data, and infrastructure together.

According to AI TIMES, 108 South Korean researchers were selected as influential scholars in the world in SearIT's 2026 report, ranking 17th among countries. The evaluation was based on citation frequency for papers published over the past 10 years, and the report was released on the 3rd. The article said researchers from various universities and research institutes in South Korea were evenly included, highlighting the breadth of the country's research base. It also emphasized that South Korea's presence was confirmed in fields such as AI, materials science, and energy, along with the distribution by domestic universities. The result is read not merely as the release of a list, but as data showing where research influence has been accumulating. Details on changes in the rankings, institutional performance, and the context behind South Korea's strengths can be found in the original article.

According to AI TIMES, Google, together with NASA's Jet Propulsion Laboratory, has unveiled a technology that uses AI to analyze satellite imagery and automatically identify the locations and spread patterns of methane emission sources. The technology, announced on the 9th, was trained on actual imagery with 3.6 million virtual methane plumes inserted and detected 84% of plumes confirmed by experts. Its key feature is that it does not stop at looking only at methane's spectral signals in hyperspectral satellite data, but also reads the shape of plumes as they spread with the wind and the surrounding terrain to distinguish real emissions from false signals. It is designed to narrow down candidate areas even when emissions are not large or when multiple plumes overlap. The researchers see the results as the foundation for a global monitoring system and have also made the database and model public so that policymakers and industry can use them for on-site inspections. With NASA's next-generation imaging spectrometers expected to further increase the volume of observations, the need for automatic analysis to supplement human interpretation is also coming into sharper focus.