Today's briefing
- Industry & Policy
Nvidia becomes Naver's third-largest shareholder with $1 billion investment
Industry & PolicyKorean outletAI TimesOct 31, 2026 - Agents
TokenEconomizer cuts coding agent token costs by up to 30.9%
AgentsMarkTechPostOct 31, 2026 - Research
Open-weight GLM 5.3 built Firefox exploits in under 48 hours
ResearchThe DecoderOct 30, 2026
Open source & papers radar
Korea Desk
What Korea’s tech press is reporting, summarized in English.
9 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Agents
Korea Deep Learning automates voice phishing response at Industrial Bank of Korea
Korea Deep Learning implemented a VLM-based automation system called Deep Agent for Finance to streamline voice phishing victim relief at Industrial Bank of Korea's consumer protection department. The system automated document recognition, classification, information extraction, and crime pattern analysis, reducing document processing time by 88% with 100% accuracy on printed text and 97% on handwritten text.
AI Times +1 moreOct 30, 2026 - Industry & Policy
Nanopozi.ai raises KRW 14.6 billion in seed funding for AI materials science
Nanopozi.ai, an AI-driven materials science startup, raised KRW 14.6 billion in seed funding from existing investors Future Play and Mashup Ventures along with new backers including Mirae Asset Venture Investment, DSC Investment, and Korea Development Bank. Total funding has exceeded KRW 15 billion.
AI TimesOct 30, 2026 - Open Source & Tools
KAIST and Newton develop carbon vertical AI with GraphRAG
KAIST's Graduate School of Green Growth and Sustainability and Newton are developing a carbon vertical AI that combines ontology, knowledge graphs, and GraphRAG to analyze global carbon markets, carbon credit transactions, and support climate policy design.
ETNewsOct 30, 2026 - Industry & Policy
Baemin's AI search grapples with non-deterministic answers
Baemin launched AI search that provides different food recommendations each time for open-ended queries like 'what should I eat when it rains.' Woowa Brothers detailed how the product team designed and validated this AI feature that lacks a single correct answer.
Byline NetworkOct 30, 2026
Industry & Policy
11 stories · 3 on the front page · 2 at the Korea DeskAll stories in this section →Stories in this issue
25 storiesCohere raises up to CAD 1.6 billion from Canadian government
Canadian AI company Cohere is receiving up to CAD 1.6 billion in investment from the Canadian government. The government is taking a direct equity stake to strengthen AI technology sovereignty and secure competitiveness in the global AI market.
OpenAI will honor robots.txt signals for AI training preferences
OpenAI announced it will honor a shared robots.txt signal to help website operators communicate their AI training preferences to participating providers.
Claude autonomously performed brain surgery on an epileptic mouse
An AI agent based on Claude autonomously selected a gene therapy, chose the brain target, and determined injection coordinates to perform surgery on an epileptic mouse. According to lead researcher Ocana-Santero, it was the first time an AI agent autonomously chose and delivered a treatment to a living mammal, though the agent refused to continue when a calibration error occurred mid-surgery.
Nvidia becomes Naver's third-largest shareholder with $1 billion investment
Nvidia will invest $1 billion in Naver through a third-party allocation, making it Naver's third-largest shareholder. The two companies set a condition requiring Naver to secure a minimum of $9 billion in computing infrastructure for its AI Factory business, which Naver fulfilled through a recent Brookfield agreement.
TokenEconomizer cuts coding agent token costs by up to 30.9%
Microsoft Research and UC Berkeley released TokenEconomizer, an open-source plugin that reduces coding agent token costs by 18.4% to 30.9% on SWE-Bench by learning from past sessions. The plugin requires no changes to the agent itself.
OpenAI updates Model Spec with five core behavioral values
OpenAI updated its Model Spec, restructuring AI behavior guidelines around five core values including faithfulness to human intent and correctability. GPT models will now draw on conversation history to infer user intent and disclose their own assumptions.
Major AI companies release security bug detection agents
Anthropic, OpenAI, Cloudflare, and Google each released AI agents to hunt security bugs in source code. All four use the same find-disprove-patch loop approach, with variations in implementation.
Open-weight GLM 5.3 built Firefox exploits in under 48 hours
Epoch AI researchers report that the open-weight model GLM 5.3 autonomously converted public security patches into working exploits for six Firefox releases in under 48 hours for about $1,000 in API fees each, without triggering safeguards from tested providers.
Anthropic's Claude detects 78% of harmful use through self-assessment
Anthropic developed a self-assessment technique where Claude evaluates conversations after responding to identify misuse risks without requiring a separate AI model. The system found large-scale fraud and illegal attempts to extract AI model information that previous safeguards missed, adding less than 2% computational overhead.
Anthropic surpasses US government as SpaceX's largest customer
Anthropic has become SpaceX's largest customer, accounting for 34.5% of SpaceX's third-quarter revenue through AI computing resources. This surpasses the US government as SpaceX's top client.
Researchers ran AI agent swarms in Minecraft and Mindustry
Researchers operated AI agent swarms in the game environments of Minecraft and Mindustry, demonstrating multi-agent coordination systems in simulated worlds.
Malicious Claude Code extension in Cursor marketplace stole data for 11 hours
A tampered Claude Code extension was live on Cursor's marketplace for approximately 11 hours, affecting up to 5,000 users. It attempted to steal user data and establish remote access before Cursor took the marketplace offline and restored it after confirming no further risk.
Korea Deep Learning automates voice phishing response at Industrial Bank of Korea
Korea Deep Learning implemented a VLM-based automation system called Deep Agent for Finance to streamline voice phishing victim relief at Industrial Bank of Korea's consumer protection department. The system automated document recognition, classification, information extraction, and crime pattern analysis, reducing document processing time by 88% with 100% accuracy on printed text and 97% on handwritten text.
Nanopozi.ai raises KRW 14.6 billion in seed funding for AI materials science
Nanopozi.ai, an AI-driven materials science startup, raised KRW 14.6 billion in seed funding from existing investors Future Play and Mashup Ventures along with new backers including Mirae Asset Venture Investment, DSC Investment, and Korea Development Bank. Total funding has exceeded KRW 15 billion.
OpenAI strengthens violence prevention and teen safety in ChatGPT
OpenAI is strengthening violence prevention in ChatGPT through expanded safeguards, earlier interventions, and a new commitment to public safety reporting.
Sterling, a 3B finance research agent, beats Claude Fable 5.1 at 11x lower cost
Samaya AI's Sterling, a post-trained Nemotron 3.5 Lightning model for financial research, scored 58.9% on FrontierFinance at $0.61 per question, outperforming Claude Fable 5.1's 55.9% at $6.89.
KAIST and Newton develop carbon vertical AI with GraphRAG
KAIST's Graduate School of Green Growth and Sustainability and Newton are developing a carbon vertical AI that combines ontology, knowledge graphs, and GraphRAG to analyze global carbon markets, carbon credit transactions, and support climate policy design.
Baemin's AI search grapples with non-deterministic answers
Baemin launched AI search that provides different food recommendations each time for open-ended queries like 'what should I eat when it rains.' Woowa Brothers detailed how the product team designed and validated this AI feature that lacks a single correct answer.
Sapper-v2, an autonomous penetration testing agent with AI commander and worker swarm
Sapper-v2 is an autonomous penetration testing legion combining an AI commander, worker swarm, and layered fact-intent graph. It won third place in Tencent's Intelligent Attack-Defense competition with 634 teams, using 1 billion tokens to outperform teams using 4.5 billion tokens over 8 hours unattended.
Google's space TPU runs Gemma 4 and writes poetry in orbit
Google's first data center satellite is operational with a TPU running a Gemma 4 model, successfully generating poetry in space.
Runware aims to mass-produce prefabricated AI data centers on assembly lines
Runware is building a factory to produce prefabricated AI data centers using assembly line methods. The company's real test will be whether it can deliver at scale.
OpenAI doubles usage allowances for ChatGPT Pro and Enterprise plans
OpenAI increased the baseline usage allowance for ChatGPT Pro and Enterprise plans by up to 2x and reversed a planned reduction for the $200 monthly tier. The move is part of OpenAI's 28-day continuous service improvement campaign.
Baseten releases joey-1, a 2B decision model at $0.01 per million tokens
Baseten released joey-1, a 2B open-weight decision model that returns calibrated probabilities instead of text. It runs at 20 ms P50 latency and costs $0.01 per million input tokens.
U.S. commerce department deems foreign robot dependence a national security risk
The U.S. secretary of commerce's Section 232 investigation concluded that dependence on robotics imports poses a national security risk.
Credit Saison achieves 954% ROI with ChatGPT Enterprise pilot
Credit Saison delivered 954% ROI from a ChatGPT Enterprise pilot, with leadership and employees helping expand the results to companywide adoption.