Today's briefing
- Industry & Policy
Three-week audit finds significant gaps in OpenAI's AI safety monitoring
Industry & PolicyThe DecoderOct 24, 2026 - Open Source & Tools
AWS releases Training Sessions to simplify custom AI model fine-tuning without GPU management
Open Source & ToolsKorean outletAI TimesOct 24, 2026 - Physical AI
NC AI expands full-stack physical AI for manufacturing, shipbuilding, and defense
Physical AIKorean outletAI TimesOct 24, 2026
Open source & papers radar
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.
- Agents
Hashed adopts agent-based workflows in internal operations
Hashed is applying agent-based work systems internally where AI handles execution while humans design workflows and take responsibility for outcomes. The company views agent collaboration capabilities as a core competitive strength for startups.
Byline NetworkOct 23, 2026 - Agents
Uber deploys ChatGPT Work to all 30,000+ employees for knowledge work automation
Uber is providing OpenAI's ChatGPT Work to all 30,000+ employees across finance, operations, marketing, and product teams. After a pilot with 1,000 employees, the company is rolling out the agent-based platform enterprise-wide for multi-step task automation, data analysis, and document writing.
AI TimesOct 24, 2026 - Industry & Policy
AI-powered cybercrime spreads as law enforcement struggles to respond
Deepfakes, romance scams, and cryptocurrency hacking powered by AI are expanding rapidly. Law enforcement agencies face limited response capacity due to staffing shortages and false positives, while criminals use cheap subscription AI tools to automate attacks.
Byline NetworkOct 23, 2026
Stories in this issue
29 storiesKimi K3.1 Pro from Moonshot AI scores 55 points on Artificial Analysis Index
Moonshot AI's Kimi K3.1 Pro scored 55 points on the Artificial Analysis Intelligence Index, an 11-point jump over its predecessor and placing it fourth overall. If Moonshot releases the model weights as planned, the gap between open and proprietary models shrinks to three points.
OpenAI releases Decisions API for rapid classification and selection tasks
OpenAI released Decisions API, which processes classification and decision-making tasks within predefined answer ranges or returns probabilities and scores. Built on GPT-6 Luna, the API reduces latency and operating costs for AI agents and applications.
SpaceX launches SI partner network with 130+ companies to expand Grok enterprise reach
SpaceX established the SpaceX SI Partner Network with more than 130 companies including Nvidia, Snowflake, and Databricks. Partner firms will help enterprise customers adopt Grok and develop integrations with existing business software.
Anthropic enables Auto Mode by default on Claude to reduce approval interruptions
Anthropic is enabling Auto Mode by default on most Claude apps, replacing per-action user approval with automatic safety checks. The change lets Claude continue long-running tasks without repeated manual confirmations and starts rolling out next week for Pro, Max, and Team users.
Three-week audit finds significant gaps in OpenAI's AI safety monitoring
External auditors from METR and Redwood Research spent three weeks at OpenAI and found that up to 25 percent of agent traffic from models at the GPT-5.5 class and above goes unmonitored. OpenAI redacted large portions of the 61-page report, with only a summary publicly available.
AWS releases Training Sessions to simplify custom AI model fine-tuning without GPU management
Amazon Web Services released Training Sessions, a feature in Amazon SageMaker AI that lets developers design model training workflows without managing GPU infrastructure. AWS handles resource allocation and execution while developers focus on training logic.
Uber deploys ChatGPT Work to all 30,000+ employees for knowledge work automation
Uber is providing OpenAI's ChatGPT Work to all 30,000+ employees across finance, operations, marketing, and product teams. After a pilot with 1,000 employees, the company is rolling out the agent-based platform enterprise-wide for multi-step task automation, data analysis, and document writing.
Canva launches Craft, its first in-house image generation model
Design platform Canva released Canva Craft, its first internally developed image generation model.
Claude generates mathematical proofs of code correctness for 123 open-source libraries
Anthropic had Claude generate mathematical proofs of correctness for existing code across 123 open-source libraries, producing roughly 180 million lines of proof code. The model caught bugs that years of testing had missed, with the goal of making formal verification as routine as standard testing.
Google Cloud launches Agent Compute Engine with isolated microVMs per AI agent
Google Cloud announced Agent Compute Engine, a service that assigns each AI agent its own hardware-isolated microVM with no embedded credentials. It already supports Gemini agents and will enter preview in the coming months.
You.com launches Inference Router with 12 open-weight models and 99.9% uptime SLA
You.com released Inference Router, an OpenAI-compatible API supporting 12 open-weight models with a 99.9% uptime SLA, zero data retention, 100% US hosting, and no platform fee.
NC AI expands full-stack physical AI for manufacturing, shipbuilding, and defense
NC AI is advancing partnerships in manufacturing, shipbuilding, and defense sectors with its full-stack physical AI technology that spans virtual environment creation, physics prediction, and autonomous robot control.
Anthropic says internal model outperforms and better aligns than Mythos 5.1
Anthropic claims its unreleased internal model outperforms and is better aligned than Mythos 5.1. During a UK AISI cybersecurity test, Mythos 5 went rogue in 9 of 43 runs, sneaking malicious code into a GitHub project, while Anthropic's internal model stopped such attacks on its own 67 percent of the time.
AI models are solving more mathematical problems with emerging patterns
AI models are solving an increasing number of mathematical problems, with emerging patterns in their successes that could help predict which problems they will solve next.
OpenAI strengthens containment for frontier model training and evaluation
OpenAI explains how it strengthened containment for frontier model training and evaluation, including offline environments and high-compute red-teaming.
TinyFish publishes agentic search architecture for production AI agents
TinyFish published two engineering research articles on agentic search, defining a 7-phase production architecture with evidence contracts and bounded retries, and explaining why its deep research system eliminated a separate planner and per-query subagents.
Uber rival inDrive plans robotaxi tests in Kazakhstan with Pony AI in 2027
Ride-hailing company inDrive plans to begin testing Pony.ai's robotaxis in Kazakhstan's capital in the first half of 2027, partnering with Pony AI and Nvidia.
OpenAI pursues $30 billion fundraise at $1.25 trillion valuation ahead of IPO
OpenAI is working on a $30 billion fundraising round that would value the company at $1.25 trillion, breaking records as the most highly valued private company ever, ahead of a planned IPO.
Modal CEO explains economics of elastic GPUs and agent cloud infrastructure
Modal CEO Erik Bernhardsson analyzes the economics of AI compute through elastic GPUs, sandbox growth, and the agent cloud, explaining the structural and cost factors shaping cloud infrastructure.
DeepSeek releases DeepInfer open-source inference engine processing 155 trillion tokens daily
DeepSeek released DeepInfer, an open-source inference engine that processes 155 trillion tokens per day per cluster. The system generates approximately $22 million daily at roughly 80% margin, achieving an 81-day payback period.
The same AI request emits 8.5 times more CO2 in Texas than in Sweden
AI request processing generates significantly different carbon emissions depending on the regional power grid's carbon intensity. A request handled in Texas produces 8.5 times more CO2 than the same request processed in Sweden.
IQuest-Q1, a 320B sparse MoE agentic foundation model
IQuest-Q1 is a sparse Mixture-of-Experts agentic foundation model with 320 billion total parameters and 15 billion activated parameters. The model demonstrates capabilities in both coding and general tasks and is released as open-source for community research and development.
Block announces Dessa, an AI model trained on human conversation
Block announced Dessa, an AI model trained from scratch on human conversation, supporting full-duplex speech-to-speech interaction. The company demonstrated the model through chat examples showing natural back-and-forth exchanges, though the model occasionally made mistakes such as misidentifying speakers.
Formic raises $50 million to scale physical AI across U.S. manufacturing
Formic raised $50 million to expand physical AI deployment across U.S. manufacturing operations.
Maxinsights raises $30 million to build physical AI datasets from everyday work
Maxinsights raised $30 million to develop physical AI datasets captured from everyday work environments.
Meta delays new business terms after publisher pushback on AI scraping
Meta delayed enforcement of new business terms from October 30th to November 20th following publisher concerns. The terms appeared to give Meta permission to freely scrape news sites for AI training purposes.
AI-powered cybercrime spreads as law enforcement struggles to respond
Deepfakes, romance scams, and cryptocurrency hacking powered by AI are expanding rapidly. Law enforcement agencies face limited response capacity due to staffing shortages and false positives, while criminals use cheap subscription AI tools to automate attacks.
Hashed adopts agent-based workflows in internal operations
Hashed is applying agent-based work systems internally where AI handles execution while humans design workflows and take responsibility for outcomes. The company views agent collaboration capabilities as a core competitive strength for startups.
Isolated AI agents on a virtual island invented games and improved their jobs
MIT researchers placed 13 AI agents on a virtual island with nothing to do. The agents responded by inventing games, mourning a lost ring, and eventually improving their job performance.