Today's briefing
- Open Source & Tools
Anthropic launches OSS Scanner, a free AI vulnerability scanner
Open Source & ToolsMarkTechPostOct 25, 2026 - Physical AI
Nvidia open-sources Flora, cutting robot video labeling cost to 72 cents per hour
Physical AIThe Decoder +1 moreOct 25, 2026 - Industry & Policy
OpenAI offers discounts up to 90 percent to win Anthropic customers
Industry & PolicyThe DecoderOct 25, 2026
Korea Desk
What Korea’s tech press is reporting, summarized in English.
8 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Industry & Policy
South Korean card firms to launch AI agent payment service next year
Shinhan Card, Hana Card, and BC Card are joining the government's "AI for All" project to launch an AI-based agent payment service in the first half of next year. The service will combine commerce and payments to create a new payment ecosystem.
ETNewsOct 25, 2026 - LLM
Tencent open-sources Hy-Nomas, a world model that edits scenes in real time
Tencent open-sourced Hy-Nomas, a world model that generates and edits video scenes based on text commands and camera movements. Users can add objects, change characters' actions, backgrounds, and art styles during generation, and the model can run on a single RTX 4090.
AI TimesOct 26, 2026 - Open Source & Tools
Liquid AI launches RIPL, a real-time inference engine
Liquid AI released RIPL, a real-time inference pipeline designed to enable AI applications in latency-sensitive tasks like search and payments. The commercial product is offered to enterprise customers to replace classical machine learning techniques with LLM-based systems.
AI TimesOct 26, 2026 - Industry & Policy
FBI investigates OpenAI engineer for economic espionage
The FBI is investigating former OpenAI engineer Abdulrahman Alfozan for economic espionage and trade secret theft, specifically whether he stole confidential information to provide to the Saudi Arabian government. He is currently charged with installing surveillance devices in a restricted conference room used by OpenAI executives.
AI TimesOct 26, 2026
Stories in this issue
20 storiesAmazon's 2026 and 2027 AI computing capacity nearly sold out
AWS computing capacity for 2026 and 2027 is nearly fully booked, with significant 2028 supply already reserved. CEO Andy Jassy said AI research labs and enterprise customers' explosive demand is straining infrastructure expansion efforts.
Anthropic co-founder Jack Clark advances self-improving AI prediction to 2027
A benchmark by P-Zero Research measuring how well AI models design and evaluate experiments on their own shows rapid improvement. Anthropic co-founder Jack Clark now predicts self-improving AI could arrive by 2027.
Hugging Face releases Carbon-X, a 14.8B DNA language model
Hugging Face released Carbon-X, a 14.8B DNA language model built on Qwen3-14B. It reads raw DNA in chat and answers in English, averaging 66.7% across 5 DNA benchmarks under the Apache 2.0 license.
Nvidia launches Nemotron 3.5 Bolt, a 4GB on-device language model
Nvidia released Nemotron 3.5 Bolt E4B 4GB, an open-source lightweight language model designed to run on 4GB of memory. It targets on-device AI systems, gaming, and robotics applications with a balance between performance and memory efficiency.
Anthropic launches OSS Scanner, a free AI vulnerability scanner
Anthropic launched OSS Scanner, a free vulnerability scanner for critical open-source projects. It runs Anthropic's Claude Mythos model against enrolled codebases and reports findings directly to maintainers without human review.
Nvidia open-sources Flora, cutting robot video labeling cost to 72 cents per hour
Nvidia's Flora data engine labels one hour of robot video for 72 cents instead of $88 with GPT-6 Astra. Built on three fine-tuned Qwen3.5 models, it catches robot action errors more reliably than competitors, though sometimes falls short on descriptive accuracy.
OpenAI offers discounts up to 90 percent to win Anthropic customers
OpenAI is poaching Anthropic's customers with steep discounts. Companies like Rippling, Replit, and DataCamp say they cut AI costs by up to 90 percent after switching from Anthropic.
Microsoft open-sources Agensh, a multi-agent harness for 1,024 agents
Microsoft Research open-sourced Agensh, a multi-agent harness with no central orchestrator. Self-organizing workers scale from 1 to 1,024 agents on ProgramBench within a 6-hour budget.
Decade-old CNN outperforms latest LLMs in skin cancer diagnosis
An Inception v3-based CNN correctly diagnosed 125 of 200 skin lesions (62%), outperforming GPT-6 Astra and Claude Fable 5.1, which each achieved 42% accuracy. The fine-tuned model was trained on public dermatology images.
Tencent open-sources Hy-Nomas, a world model that edits scenes in real time
Tencent open-sourced Hy-Nomas, a world model that generates and edits video scenes based on text commands and camera movements. Users can add objects, change characters' actions, backgrounds, and art styles during generation, and the model can run on a single RTX 4090.
Liquid AI launches RIPL, a real-time inference engine
Liquid AI released RIPL, a real-time inference pipeline designed to enable AI applications in latency-sensitive tasks like search and payments. The commercial product is offered to enterprise customers to replace classical machine learning techniques with LLM-based systems.
FBI investigates OpenAI engineer for economic espionage
The FBI is investigating former OpenAI engineer Abdulrahman Alfozan for economic espionage and trade secret theft, specifically whether he stole confidential information to provide to the Saudi Arabian government. He is currently charged with installing surveillance devices in a restricted conference room used by OpenAI executives.
TokenGate accelerates long-context prefill by up to 11.3x
Researchers from Stanford and UC Berkeley released TokenGate, a learned token pruning module for LLMs. It cuts long-context prefill time-to-first-token by 1.4x to 11.3x and peak memory by 1.5x to 3.4x without accuracy loss.
Anthropic recommends consolidating sub-agents into single agents
Anthropic recommends that companies evaluate consolidating multiple sub-agents into a single agent that loads context on demand. In one architecture comparison, the single-agent setup ran about twice as fast, and Anthropic published a migration kit on GitHub.
OpenMobile-2 proves open-source smartphone agents can match proprietary models
OpenMobile-2 demonstrates that open-source smartphone agents trained on diverse apps can rival proprietary models. The model was trained on more than 110 applications, including 35 app replicas, and scores highest among open agents on real Android devices.
Pennsylvania requires labels on AI-generated ads
Pennsylvania Governor Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act into law. It requires any ad in still image, video, audio, or text format that has been significantly modified or generated by artificial intelligence to include a label in the same format.
MIT finds AI agents spontaneously invent play and discover unknown techniques
MIT researchers found that AI agents placed on a virtual island without specific tasks spontaneously created games like basketball and bowling, and discovered object-throwing techniques unknown to the simulator's developers. These self-directed explorations improved the agents' ability to perform subsequent tasks.
GPT-6 Astra decodes messages Claude hid in images with 88% accuracy
In an experiment, Claude Fable 5.1 hid messages in images by encoding text as star positions and building heights. GPT-6 Astra decoded 44 of 50 hidden messages (88%) without receiving the encoding rules, analyzing visual patterns to restore the text.
Kimi K3.1 achieves Pareto-optimal position on cost-efficiency benchmark
Kimi K3.1 Pro scored 55 points on the Artificial Analysis Intelligence Index, well above the median of 27 points for models in its price range. The evaluation cost is $2.11 per task.
South Korean card firms to launch AI agent payment service next year
Shinhan Card, Hana Card, and BC Card are joining the government's "AI for All" project to launch an AI-based agent payment service in the first half of next year. The service will combine commerce and payments to create a new payment ecosystem.