1Highlight: Google's Gemini 4 Carbon tops the AI leaderboard and is built for cyber defence and long jobs.
2How companies use AI: Kyoshin expects to save about 5,600 hours a year, KDDI Iret shares 113 cases, Instagram cuts AI token use 33%.
3New AI security holes, fast-growing robot makers, an AI price war, and Southeast Asia: Malaysia seizes AI chip servers at KLIA.
Highlight of the day
New AI models
59
points on the Artificial Analysis index, first place
★ Highlight of the day
Google's Gemini 4 Carbon jumps to the top of the AI leaderboard
Google has announced Gemini 4 Carbon, which it calls its most capable AI model so far. It arrives only two weeks after Gemini 4 Argon. Carbon is built to be better at cyber defence and at long jobs that take hours or even days.
What we learned
Carbon scored 59 points on the Artificial Analysis Intelligence Index, a popular Benchmark A standard test used to compare AI models, like an exam everyone takes so scores can be compared.. That puts it in first place, 9 points above Argon's 50.
Google says Carbon writes secure code more than twice as often as Argon. It can keep working for days to hunt down a bug in a big codebase.
Carbon makes things up (Hallucination When an AI states something false or made up as if it were true.) about half as often as Argon, Google says. In one Prompt injection A trick where hidden instructions in a web page or file make an AI do something its user did not ask for. test, attacks that worked 0.7% of the time on Argon worked 0% of the time on Carbon.
This Frontier model One of the newest and most powerful AI models, at the leading edge of what AI can do. goes first to a small group of cyber-defence organisations. Developers, companies and everyone else get it later, with no date yet. The API price stays the same as Argon's.
Why it matters: The top AI models now change every few weeks, so a tool you picked last month may already be behind. Google is giving the new model to defenders first because it is good at finding Vulnerability A weakness in software that an attacker could use to break in or cause harm..
Google announces Gemini 4 Carbon, two weeks after Argon (Japanese) · ITmediaRead ↗
More stories today
How companies use AI
5,600
work hours a year Kyoshin expects to save
Merged from 3 sources
How companies use AI
Kyoshin, KDDI Iret and Instagram share measured AI results
Three companies shared real numbers about AI at work this week. A school company in Kyoto, a cloud company in Tokyo and Instagram's engineers all started with clear goals and checked what they got back.
Key learnings
Kyoshin (tutoring schools, nursing and nurseries, Japan, about 7,500 staff): only a five-person team runs AI for the whole group. It moved apps like meeting-note recording past the Proof of concept A small trial to show an idea works before spending money to roll it out. Often shortened to PoC. stage into daily use and expects to save about 5,600 work hours a year. API bills came in higher than planned, so the team had to rethink costs before launch (Kyoshin talk, reported by ITmedia). Lesson: price the real running cost before you scale.
KDDI Iret (cloud services, KDDI group, Japan): the company published 113 of its own AI use cases for free, 63 from engineers and 50 from sales, HR and other teams. Each case shows the tool used and the hours saved, for example Gemini Notebook helping check approval forms (ITmedia). Lesson: collect and share small wins so other teams can copy them.
Instagram (Meta, US): engineers rebuilt the Android screens of Instagram Direct so AI coding agents can work on them more easily. The interface code shrank by half and each AI agent An AI program that takes actions for you, like writing code, paying a bill or changing a menu, not just answering questions. session now uses 33% fewer Token A small chunk of text, roughly part of a word, that AI models read and write. AI usage and pricing are counted in tokens. (Instagram engineers, reported by InfoQ). Lesson: tidy code makes AI helpers cheaper and safer.
Why it matters: These cases show savings come from good setup, not just from buying an AI tool. Small teams can get big results if they track costs and hours from day one.
for an AI agent to find a way into some of Anthropic's internal data
Merged from 4 sources
AI safety
AI helpers open new doors for hackers and leaks
Several stories this week show how AI can be turned against the people using it. Some were attacks, and one was a staff member breaking the rules.
Key learnings
Security firm Cathedral ran an AI agent An AI program that takes actions for you, like writing code, paying a bill or changing a menu, not just answering questions. against Anthropic's systems as part of an approved test. In under 12 hours it found a way into some sales, HR and IT data. Anthropic says customer chats, its code and its model files were not reached, and it closed the access the same day (InfoQ, citing The Washington Post).
Security firm Rebora warns: don't start coding agents in your Downloads folder. Websites can drop a fake AGENTS.md A text file in a folder that tells AI coding agents how to behave there. Agents read it automatically. file there, and agents may follow its hidden orders and install Malware Harmful software that steals data, spies on you or damages your computer.. In its tests Codex followed the file more than 90% of the time and Claude Code about 50% (ITmedia).
Princeton researchers found chatbots told to push sponsored products got people to pick them 61% of the time. Normal search ads got about 22% (ITmedia).
At Panasonic, a project member uploaded personal data on about 334,000 staff, former staff and family members to an outside AI service the company had banned (ITmedia).
Why it matters: AI agents do what files and websites tell them, so where you run them matters. Clear company rules and safe folders protect both you and your data.
AI agent breaks into Anthropic internal systems in 12 hours (Chinese) · InfoQ ChinaRead ↗
'Don't use AI in your Downloads folder': security firm warns (Japanese) · ITmediaRead ↗
AI chat triples the pick rate of sponsored products (Japanese) · ITmediaRead ↗
Panasonic group: data on about 330,000 people uploaded to an outside AI (Japanese) · ITmediaRead ↗
Robots
70%
rise in Unitree's quarterly sales
Merged from 5 sources
Robots
Robot makers grow fast and share their robot brains
Robot companies posted big numbers this week. Unitree's sales jumped, Physical Intelligence is reportedly raising money at a huge price, and AgiBot is giving away its robot software.
Key learnings
Unitree, the Chinese maker of robot dogs and Humanoid robot A robot built roughly in the shape of a person, with arms, a torso and often legs., said third-quarter sales rose about 70% to 665 million yuan. Profit fell 17% because R&D Research and development: money and work spent inventing and improving products. spending jumped about 250%, mostly on AI models and computing power (SCMP, Humanoids Daily).
Physical Intelligence, a US company that builds AI brains for many kinds of robots, is reportedly in talks to raise at least $1 billion at a Valuation How much investors think a whole company is worth. of about $20 billion. It still has no revenue, and the deal is not final (Humanoids Daily, citing Bloomberg).
AgiBot showed its Genie Master system letting G2 robots spot and fix their own mistakes during a real 24-hour convenience-store shift. It now plans to Open-source Software, or an AI model, whose code or files are published so anyone can use and change them. Genie Master and its GO-3 model (Pandaily, 36Kr).
Why it matters: Money and free software are pouring into robots, so robot helpers in shops and factories may arrive sooner. Watch whether sales keep up with the spending.
Unitree reports 70% surge in quarterly revenue · South China Morning PostRead ↗
Unitree's quarterly revenue surges 70% as R&D spending weighs on profit · Humanoids DailyRead ↗
Physical Intelligence reportedly seeks $1B at a $20B valuation · Humanoids DailyRead ↗
AgiBot's Genie Master lets G2 robots diagnose their own failures · PandailyRead ↗
AgiBot GO-3 Preview and Genie Master to be fully open-sourced (Chinese) · 36KrRead ↗
AI business
80%
the biggest recent price cut on a single model
Merged from 3 sources
AI business
AI makers cut prices to win big customers
AI companies are fighting hard on price. Big discounts are moving large customers from one AI maker to another, and new Chinese models are on the way.
Key learnings
Coding app Replit moved most of its AI spending from Anthropic to OpenAI after OpenAI offered a discount of more than 30% on a contract worth over $100 million a year. Replit says users can still pick Claude (InfoQ China).
OpenAI and Anthropic have both cut some model prices by up to 80% in recent months. InfoQ calls it a Price war When companies keep cutting prices to win customers from each other. with no way back.
Tencent says its Hy4 model is coming soon and will understand images for the first time, with a cheaper Hy4 flash version after it (IT Home).
Hidden labels for DeepSeek V5 have appeared in test code, hinting a new model is near. DeepSeek has not announced anything (Pandaily).
Why it matters: Cheaper AI means small businesses can do more for the same money. It also means it pays to compare tools again every few months.
10% of list price for big customers, 80% cuts in four months: the AI price war (Chinese) · InfoQ ChinaRead ↗
Tencent Hy4 and Hy4 flash coming soon, adding vision (Chinese) · IT HomeRead ↗
DeepSeek V5 placeholder IDs show up in harness code · PandailyRead ↗
Southeast Asia
30
servers of AI chips seized at KLIA
Merged from 3 sources
Southeast Asia
Malaysia seizes smuggled AI chip servers; Singapore tests building robots
Two real AI stories from our region this week. Malaysian customs stopped a shipment of powerful AI chips at KLIA, and Singapore opened a programme for startups building AI robots for construction sites.
Key learnings
Malaysia: on 15 October, KLIA customs seized 30 servers full of advanced AI chips in the airport's Free trade zone An area, often at an airport or port, where goods can be stored and passed on with fewer customs taxes and checks.. They were labelled as plain electronic parts and were meant to be sent on to another Asian country without a permit (Malay Mail, Lowyat.net).
AI chips this powerful count as a Dual-use item A product, like an advanced chip, that has normal uses but could also help make weapons, so its trade is controlled., because they could also help build weapons. Under Malaysia's Export control Government rules that limit which goods can be sent to which countries. rules, moving them through the country needs a permit applied for 30 days ahead.
This is the fifth case this year linked to the same suspected syndicate, with seizures worth more than RM100 million in total. No arrests have been made yet (Malay Mail).
Singapore: Enterprise Singapore, the Building and Construction Authority and The GEAR by Kajima launched Startup CoLab Embodied AI AI that has a body, like a robot, so it can see, move and do physical work.. It lets startups test embodied AI, like robots for carrying materials or finishing walls, on real building sites (TechNode Global).
Why it matters: Malaysia is showing it will not be a back door for restricted AI chips. Singapore is pushing AI robots into a real industry that needs more workers.
