Today's briefing
- Physical AI
Google releases Gemini 4 Argon for robotics with enhanced physical reasoning
Physical AIKorean outletAI TimesOct 15, 2026 - Industry & Policy
Nvidia pledges $2 billion in European investment and $250 million in UK AI safety compute
Industry & PolicyThe DecoderOct 14, 2026 - Agents
SentinelOne cuts malware analysis from weeks to hours with OpenAI Daybreak
AgentsOpenAI News +1 moreOct 14, 2026
Open source & papers radar
Korea Desk
What Korea’s tech press is reporting, summarized in English.
11 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Physical AI
Physical AI commercialization hinges on frontier models and inference infrastructure
As physical AI moves toward commercialization, frontier model development and inference efficiency have become critical. FriendliAI's CEO said that in the agent era, inference infrastructure matters as much as the frontier models that serve as robot brains.
ETNewsOct 14, 2026 - Agents
Kakao emphasizes full-stack optimization to power AI agents for 50 million KakaoTalk users
Kakao's Kanana model team stressed that intelligent models alone are insufficient for an AI agent serving 50 million users around the clock. The company outlined three strategies—full-stack optimization, multimodal capabilities, and agentic AI—plus memory and safe action execution as key requirements.
Byline Network +2 moreOct 14, 2026 - Physical AI
MakinaRocks joins KEIT project to automate automotive forging with AI
MakinaRocks was selected as a collaborative research organization for South Korea's KEIT industrial technology development project to automate hot forging processes in automotive parts using AI. The project runs through December 2029 and allocates KRW 2.4 billion to MakinaRocks.
ETNewsOct 14, 2026 - Agents
Public sector AI platform adoption expands with LG CNS and Samsung SDS registrations
LG CNS's agentic AI platform AgenticWorks and Samsung SDS's enterprise AI platform were registered on the Digital Service Utilization Support System, broadening options for public agencies to adopt AI platforms and connect internal data to business systems.
ETNewsOct 14, 2026
Stories in this issue
29 storiesOpenAI racing to finish flagship model using looped transformers for math and coding
OpenAI is finishing its next flagship model, which uses a looped transformer technique to boost math and coding performance. The technique makes safety monitoring harder, and OpenAI is focusing on alignment work before release. Employees say one or two new models will likely ship by year's end.
Anthropic opens Singapore office and signs AI safety MOU with government
Anthropic opened its fifth Asia-Pacific office in Singapore and signed a Memorandum of Understanding with the Singapore Ministry of Digital Development and Information on AI safety, security, and responsible adoption. Claude is already powering services for millions of users at GovTech Singapore, Grab, and DBS.
Anthropic warns small AI models approaching capability to autonomously replicate via IoT devices
Anthropic warned that small language models with 30 billion parameters or fewer are approaching the ability to autonomously infect, replicate, and propagate across internet-connected IoT devices. These AI worms lack a central server, making them harder to block and remove than traditional malware.
Anthropic strengthens defenses against illegal model distillation, reducing data extraction by 90-99%
Anthropic significantly strengthened Claude's defenses against illegal distillation attacks that steal the model's reasoning capabilities. New security measures reduced the amount of inference data attackers can collect by 90 to 99 percent compared to previous levels.
Google releases Gemini 4 Argon for robotics with enhanced physical reasoning
Google released Gemini 4 Argon for Robotics to help robots understand physical environments and improve their own task performance. The model reasons jointly about video, 3D space, and physics laws to enable complex manipulation and boost task efficiency.
Nvidia pledges $2 billion in European investment and $250 million in UK AI safety compute
Nvidia CEO Jensen Huang announced $2 billion in European investments and stakes in French startups including Doctolib and Entalpic. The UK will receive $250 million in compute for AI agent safety and security research.
SentinelOne cuts malware analysis from weeks to hours with OpenAI Daybreak
SentinelOne uses OpenAI Daybreak to cut weeks of malware analysis down to hours, freeing capacity to investigate more threats.
Google releases Gemini 4 Argon API with 1 million token output limit
Google made Gemini 4 Argon generally available on the Gemini API, AI Studio, and Gemini app. The sparse MoE model writes up to 1 million output tokens and leads the Vals Index at 68.90%.
Google DeepMind and Singapore jointly develop on-device AI for scam detection
Google DeepMind and GovTech Singapore developed a lightweight on-device Gemma model to intercept conversational scams in private chats. The system detects fraud without sending data to the cloud.
ASML's strong orders ease fears AI spending boom is cooling
ASML reported strong order intake, easing concerns that the AI infrastructure investment boom may be losing momentum.
Last Secure Code Benchmark evaluates secure code generation by AI agents
The Last Secure Code Benchmark comprises 450 tasks built from 150 real vulnerabilities across seven languages and 21 of the 2025 CWE Top 25. Even Claude Fable 5, the strongest agent tested, produces code that is both functional and secure for only 25.1 percent of tasks.
Google says self-improving AI is nowhere close to reality
Google engineer Haydn Belfield says DeepMind has found no measurable acceleration of AI research driven by AI itself. Although AI now writes over 75 percent of production code, 88 percent of changes still require human oversight, and the General Capability Index shows steady growth with no sudden speedup.
OpenAI focuses on safety for next-generation model using Loop Transformer design
OpenAI is working to safely release a next-generation model that uses Loop Transformer architecture, which makes it harder to monitor the AI's reasoning process. The company is concentrating on alignment work and security enhancements to ensure the model and test environment are safe for leadership to publicly release.
Apate sends AI agents undercover to gather fraud intelligence
Apate.ai deploys AI agents to get scammed deliberately in order to collect fraud intelligence and help combat scam networks.
Z.ai releases GLM-Legal, a law-specialized model that beats Claude Fable 5
Zhipu AI released GLM-Legal, a model fine-tuned from GLM-5.3-Flash for legal document analysis and task completion. It scored higher than Anthropic's Claude Fable 5 and Moonshot AI's Kimi K3 in legal agent performance evaluation.
Artemis Security's Orion-1 fills gap in AI-powered cyber defense
Artemis Security released Orion-1, a model purpose-built for cyber defense that beats Claude Opus 5.5 and GPT-6 Sol on the company's benchmark. It addresses a growing need as general-purpose frontier models tighten restrictions on cyber capabilities, blocking legitimate defenders.
Physical AI commercialization hinges on frontier models and inference infrastructure
As physical AI moves toward commercialization, frontier model development and inference efficiency have become critical. FriendliAI's CEO said that in the agent era, inference infrastructure matters as much as the frontier models that serve as robot brains.
Liquid AI releases LFM2.5-24B-A2B, a 2B-active MoE model for on-device agents
Liquid AI released LFM2.5-24B-A2B, an open-weight 24B Mixture-of-Experts reasoning model with 2B active parameters. On an Apple M5 Max, it scores 48.09 on LiveClaw and decodes 184 tokens per second, targeting local agent deployment.
OpenAI analysis shows agentic AI adoption growing across EU enterprises
New analysis from OpenAI demonstrates how agentic AI adoption is expanding among EU enterprises and small businesses, and identifies where deeper use could unlock additional value.
Meta launches Muse Wonder and Muse Wonder Full, 1M-context agent models
Meta released Muse Wonder and Muse Wonder Full, agent models supporting 1 million tokens of context. Muse Wonder costs $4 to $20 per million tokens on Meta Model API, while Wonder Full leads on DeepSWE and SWE-Marathon benchmarks.
Anthropic strengthens election safeguards ahead of 2026 US midterms
Anthropic analyzed 250,000 Claude.ai conversations over one week, finding that about 5.75 percent touch on politics or policy, while 0.51 percent concern elections or voting. Requests asking who or what to vote for appeared in only 0.01 percent of US conversations.
Microsoft develops machine learning tool for accurate two- to six-week weather forecasts
Microsoft published findings in Nature on a cost-effective machine learning approach to improve weather forecasts two to six weeks in advance, outperforming traditional physics-based models and recent AI approaches. The tool is now available as open-source software.
Android deploys AI to detect scammers in Google Messages
Android is rolling out enhanced protection in Google Messages using cloud-based Gemini AI models to evaluate potential scammer intentions in conversations with unrecognized senders, strengthening defenses against mobile fraud.
Kakao emphasizes full-stack optimization to power AI agents for 50 million KakaoTalk users
Kakao's Kanana model team stressed that intelligent models alone are insufficient for an AI agent serving 50 million users around the clock. The company outlined three strategies—full-stack optimization, multimodal capabilities, and agentic AI—plus memory and safe action execution as key requirements.
NetApp unveils Novus, an AI storage platform to eliminate GPU data bottlenecks
NetApp introduced Novus, a next-generation storage platform designed to reduce data bottlenecks when supplying GPUs in AI factories. It enhances data utilization and security for agentic AI and strengthens hybrid multicloud management.
MakinaRocks joins KEIT project to automate automotive forging with AI
MakinaRocks was selected as a collaborative research organization for South Korea's KEIT industrial technology development project to automate hot forging processes in automotive parts using AI. The project runs through December 2029 and allocates KRW 2.4 billion to MakinaRocks.
maum.ai pursues regulatory exemption for quadruped robot sidewalk patrols
maum.ai received review of its regulatory exemption application to enable quadruped robots to patrol pedestrian sidewalks. Current safety certification requirements focus on wheeled robots, creating a gap for legged robot deployment.
Public sector AI platform adoption expands with LG CNS and Samsung SDS registrations
LG CNS's agentic AI platform AgenticWorks and Samsung SDS's enterprise AI platform were registered on the Digital Service Utilization Support System, broadening options for public agencies to adopt AI platforms and connect internal data to business systems.
Dyna-2 achieves hour-long autonomy for general-purpose physical agents
Dyna Robotics unveiled Dyna-2, a general-purpose physical agent built on World Action Models that achieves hour-long autonomous operation.