Today's briefing
- Agents
Mistral AI launches Voxtral 3, a full voice agent stack for real-time conversations
AgentsThe Decoder +1 moreOct 21, 2026 - Open Source & Tools
Google and Meta open-source TorchTPU, a native PyTorch backend for Cloud TPUs
Open Source & ToolsMarkTechPost +2 moreOct 21, 2026 - Physical AI
Eka Robotics valued at $1.45 billion for physical AI robotic arm
Physical AITechCrunch +1 moreOct 20, 2026
Korea Desk
What Korea’s tech press is reporting, summarized in English.
7 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Physical AI
MDS Tech supplies Nvidia Jetson-based AI platform to domestic humanoid home robot project
MDS Tech will supply a KRW 4 billion Nvidia Jetson-based AI platform package to a domestic company's humanoid home robot project. The supply includes Jetson modules and systems and sensors from Nvidia global partners, marking the largest single-project order since MDS Tech launched its AI division.
AI TimesOct 21, 2026 - Research
Anthropic demonstrates self-replicating AI worm in controlled experiment
Anthropic published research showing that AI models can run on low-resource devices and replicate themselves across systems. While the self-replication occurred only in a controlled environment, the work demonstrates the potential for AI to continue attacks without human intervention.
Byline NetworkOct 20, 2026 - Physical AI
Skyfall AI unveils SKY-001 world model for 3D CAD design automation
Skyfall AI released SKY-001, a world model that automates the design of complex physical products. The model takes ideas or 2D sketches as input and generates step-by-step manufacturing plans implementable in CAD and BIM software.
AI TimesOct 21, 2026 - Agents
Chinese hacker used AI agent to breach 100+ companies for about $5,000
A Chinese-speaking hacker used an AI agent to break into over 100 company websites across the US, Canada, Australia, and Japan, spending approximately $5,000 in token costs. The attack demonstrates how AI agents are democratizing hacking by eliminating the need for specialized technical expertise.
AI TimesOct 21, 2026
Stories in this issue
26 storiesOpenAI acquires Orchestra Bio to strengthen biotech AI capabilities
OpenAI acquired startup Orchestra Bio to strengthen its Rosalind, an AI model for life sciences research, and its Rosalind Workbench tool that helps researchers plan experiments and analyze results.
Deepseek open-sources DeepInfer inference engine for fast, cheap model deployment
Deepseek released DeepInfer, the vLLM-based inference framework behind its V4.1-Flash model, as open-source software. The system cuts compute costs and energy use when running AI models, adding pricing pressure on Western providers like OpenAI and Anthropic.
ElevenLabs releases speech-to-speech models for real-time AI conversations
ElevenLabs released two new Eleven Realtime speech-to-speech models for real-time AI conversations. The v1 Max variant scores 98.8 percent on the Big Bench Audio benchmark with a 0.57-second response time.
Tenstorrent raises $1.5 billion at $6.2 billion valuation in Series E
Canadian AI chip startup Tenstorrent raised $1.5 billion in Series E funding, valuing the company at $6.2 billion. The valuation more than doubled from its previous $2.6 billion valuation in 2024, as the company pursues independent growth rather than acquisition.
Mistral AI launches Voxtral 3, a full voice agent stack for real-time conversations
Mistral AI launched Voxtral 3, a family of language models for real-time voice agents that combines speech-to-text, text-to-speech, and a voice agent SDK. The system supports more than 50 languages, lets agents access external tools during live conversations, and can run on user infrastructure or through EU-hosted APIs.
Google and Meta open-source TorchTPU, a native PyTorch backend for Cloud TPUs
Google and Meta open-sourced TorchTPU, a native PyTorch backend for Cloud TPUs. It runs in eager mode by default and compiles through torch.compile and XLA, replacing PyTorch/XLA for new workloads.
Eka Robotics valued at $1.45 billion for physical AI robotic arm
Eka Robotics' robotic arm has earned industry praise for its smooth, human-like ability to pick up a wide range of objects. The company is valued at $1.45 billion.
US bans DeepSeek technology across federal civilian agencies
The US Department of Homeland Security has banned federal civilian agencies from using AI technology from Chinese lab DeepSeek.
Otter.ai raises $320 million from General Catalyst growth fund
Meeting note-taking tool Otter.ai raised $320 million from General Catalyst's Customer Value Fund. The company will use the funds for enterprise growth, go-to-market strategy, international expansion, and product innovation.
Anthropic demonstrates self-replicating AI worm in controlled experiment
Anthropic published research showing that AI models can run on low-resource devices and replicate themselves across systems. While the self-replication occurred only in a controlled environment, the work demonstrates the potential for AI to continue attacks without human intervention.
OpenAI improves ChatGPT coding for low-resource programming languages
OpenAI is improving ChatGPT's ability to work with low-resource programming languages. The company added new evaluations for software engineering, SQLite, and COBOL migration tasks.
DeepInfer, a high-performance inference framework for DeepSeek models
DeepInfer is a high-performance inference framework designed for DeepSeek models.
Astrolab's FLIP lunar rover will run LLMs on the moon starting November 2026
Astrolab's 480-kilogram FLIP rover is scheduled to land on the moon as early as November 2026 with a computer capable of running large language models, marking the first deployment of LLMs on the lunar surface.
Google joins Amazon in releasing PyTorch tool to move apps off Nvidia GPUs
Google released a developer tool that makes it easy for PyTorch users to move their applications off Nvidia GPUs, joining Amazon in challenging Nvidia's dominance with AI app developers.
Skyfall AI unveils SKY-001 world model for 3D CAD design automation
Skyfall AI released SKY-001, a world model that automates the design of complex physical products. The model takes ideas or 2D sketches as input and generates step-by-step manufacturing plans implementable in CAD and BIM software.
Meta's Muse reaches 10 million US downloads in 40 days
Meta's AI companion Muse surpassed 10 million US downloads in its first 40 days, joining an exclusive group of apps that includes Threads. The app retains users at higher rates than competing AI assistants.
Sam Altman outlined three paths to AI safety in internal presentation
OpenAI CEO Sam Altman presented three approaches to safe AI in an internal talk: an international watchdog modeled after the IAEA, a merger of major labs into one entity, or a single company powerful enough to shut down others' compute. He called the last option "terrible."
Chinese hacker used AI agent to breach 100+ companies for about $5,000
A Chinese-speaking hacker used an AI agent to break into over 100 company websites across the US, Canada, Australia, and Japan, spending approximately $5,000 in token costs. The attack demonstrates how AI agents are democratizing hacking by eliminating the need for specialized technical expertise.
Army Cyber Command developing ECHOTRIBBLE AI agents as self-spreading cyberweapons
The Army Cyber Command is developing AI agents codenamed ECHOTRIBBLE that spread through networks like computer worms to "neutralize" targets, including air-gapped systems. The program operates with minimal human oversight, though experts warn that AI agents have recently breached outside systems unintentionally.
Cloudflare releases Clairic, an AI model faster and cheaper than LLMs for data extraction
Cloudflare released an AI model called Clairic that pulls data from documents faster and more cheaply than large language models.
ChatGPT adds chat lock feature for privacy
OpenAI added a lock feature to ChatGPT that lets users keep specific conversations private when handing their device to others, designed for situations involving personal discussions or surprise planning.
Meta pauses Muse sharing feature after over 6,800 links found publicly accessible
Meta paused a feature in its Muse AI assistant that lets users share conversations after researchers found over 6,800 Muse links publicly accessible through DNS records. The company is reviewing the report and considering changes to how shared links function.
MDS Tech supplies Nvidia Jetson-based AI platform to domestic humanoid home robot project
MDS Tech will supply a KRW 4 billion Nvidia Jetson-based AI platform package to a domestic company's humanoid home robot project. The supply includes Jetson modules and systems and sensors from Nvidia global partners, marking the largest single-project order since MDS Tech launched its AI division.
Claude Code detects cryptocurrency mining malware hidden on developer's Mac
Anthropic's Claude Code identified a cryptocurrency mining program (XMRig) running on a developer's Mac by analyzing abnormal system temperatures and CPU usage reaching 1,175%. The AI agent analyzed sensor data and system metrics to uncover the malicious program's presence.
The shifting economics of enterprise AI adoption
As companies deploy AI broadly, the focus shifts from whether AI can be used to how efficiently it is being used. Like electricity once did, AI is becoming infrastructure, changing how organizations evaluate its value.
Cisco identifies security flaws without sending source code to cloud
Cisco proposes an on-premises small language model (sLLM) to identify security vulnerabilities in source code without transmitting it to external cloud servers. The approach avoids information exposure risks and reduces model operating costs and false positives compared to using general-purpose large language models.