1Highlight: ChatGPT Wallet lets OpenAI's dots agents buy things for you, with a Face ID or fingerprint check on every purchase.
2Malaysia: MITI says it never cleared Megaspeed or Speedmatrix; Singapore barred ex-CEO Alice Huang from re-entering.
3Also: Microsoft's 18-hour outage, OpenAI breaks up its robotics team, and Google's yes-or-no Gemini Decision 1.
Highlight of the day
AI agents
1
security check needed for every purchase
Merged from 2 sources
★ Highlight of the day
ChatGPT Wallet lets OpenAI's agents pay for things after you approve
OpenAI launched ChatGPT Wallet on 13 November (US time), so its ChatGPT AI agent An AI program that can take actions for you, like writing code, filling forms or moving items, not just answering questions., called dots, can finish purchases for you. It starts with US ChatGPT Pro subscribers aged 18 and over who have an eligible US-issued Visa, Mastercard or American Express card.
What we learned
How it works: you add a card once. Before each purchase you see the items, the total and the card, then approve with a check such as Face ID or a fingerprint. OpenAI says each approval leaves a record of what you allowed the agent to buy.
The agent never sees your real card number. After you approve, the card network issues a Payment token A stand-in card number issued for a purchase, so a shop or app never sees your real card details., a stand-in card number, so neither the dot nor the shop gets your actual details. OpenAI says card rewards and protections for eligible disputes still apply.
Shops may still block or decline agent purchases. OpenAI published a Commerce Partner Playbook for retailers and is working with payment firms such as Stripe, Adyen and Checkout.com and fraud firms such as Forter and Signifyd. TechCrunch notes Amazon bans outside AI agents from its store.
Later, OpenAI expects to let users set spending limits, product categories or time periods so the agent can buy within pre-approved rules. That is a plan, not a feature yet.
Why it matters: This is one of the first agent payment systems built with all three big US card networks, and for now a person still approves every purchase. It is not available in Malaysia yet; if it arrives, keep the per-purchase check and read each total before you approve.
ChatGPT's new wallet lets AI agents shop and check out for you · TechCrunchRead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
More stories today
Singapore & Malaysia
917
servers in Singapore's latest money-laundering charges
Merged from 4 sources
Singapore & Malaysia
Singapore bars ex-Megaspeed boss as Malaysia says it never cleared Megaspeed
Singapore's Ministry of Home Affairs said on 12 November that Alice Huang, former chief executive of Singapore-based Nvidia Cloud computing Renting computers in someone else's data centre over the internet instead of buying your own. partner Megaspeed, got a stern warning and is barred from re-entering Singapore. On 13 November Malaysia's trade ministry MITI said it had never cleared Megaspeed or its Malaysian affiliate Speedmatrix.
Key learnings
Singapore: after consulting the Attorney-General's Chambers, police found too little evidence that Huang committed an offence, so she was not charged. They say she knowingly gave inaccurate information to play down her role. She was not investigated for spying; police held her passport from June 2025 until 26 February 2026.
Why Malaysia is involved: the ministry says servers bought by companies under investigation went through Singapore to Malaysia without entering Singapore's market. Police are checking whether they were then sent on somewhere else, and say server maker Supermicro has given only limited help despite requests to US authorities.
Malaysia: MITI says its earlier Data centre A large building full of computers that store data and run online services and AI, and that uses a lot of power and water. visits were routine checks, not approval. With the Royal Malaysian Customs Department it is checking whether chips brought in for Speedmatrix were Re-export Sending imported goods on to another country instead of using them where they first arrived.. Bringing AI chips in needs no permit, but exporting or re-exporting high-end US AI chips has needed a Strategic Trade Act Malaysia's 2010 law requiring permits to export or move sensitive goods and technology, now including high-end US AI chips. permit since July 2025.
In a separate Nvidia-chip case, Singapore charged a fifth person, Shi Bingzhe, on 13 November with 41 Money laundering Hiding where money or goods obtained through crime came from, for example by passing them through companies. counts over 917 servers bought from Aperia International, worth more than US$100 million (about RM408 million), CNA reports. Megaspeed's chief executive says it is cooperating with US inspections.
Why it matters: Both governments are trying to show they are not Backdoor A hidden way into a computer or system that lets an attacker get back in later without the owner knowing. for restricted US chips. Data centre operators and GPU A powerful computer chip, originally for graphics, that is now the main hardware used to train and run AI. tenants in Malaysia should expect more questions about where every imported server ends up, and should keep permits and shipping records in order.
Alice Huang barred from re-entering Singapore; insufficient evidence she committed offences in Megaspeed probe: MHA · CNARead ↗
MITI: Megaspeed and Speedmatrix was never given the all-clear in Malaysia · Lowyat.netRead ↗
Malaysia checking chip re-export; Singapore-based Megaspeed not yet cleared · TechNode GlobalRead ↗
Fifth person charged in Singapore over alleged fraud conspiracy linked to Nvidia chips · CNARead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
Cloud outage
18 hours
of downtime before traffic returned (Kentik)
Merged from 4 sources
Cloud outage
Microsoft outage took Outlook, Teams and Copilot offline for 18 hours
A Microsoft 365 outage that began early on 13 November (Malaysia time) knocked out Outlook, Teams, Copilot and other Cloud computing Renting computers in someone else's data centre over the internet instead of buying your own. services around the world. Microsoft says servers were failing Health check An automatic test a system runs to see if a server is working; servers that fail are taken out of service. and that it has found no evidence of malicious activity.
Key learnings
What broke: Microsoft said a core process on affected servers became overwhelmed, so servers failed health checks and dropped out; some failed again once traffic returned. It extended email timeouts to 48 hours so delayed mail would not be lost, and promised a full Post-incident report A company's write-up after an outage or breach explaining what went wrong and how it will stop a repeat. on the Root access Full control of a computer, able to read, change or delete anything on it. cause.
Who was hit: Malay Mail, citing AsiaOne, says OpenAI and Anthropic reported problems and Singapore's GovTech A government unit that builds digital tools for public services. warned email replies could be delayed. Ars Technica, citing NBC, says Japan's government IT systems and the Philippines' Supreme Court were disrupted.
AI tools broke too: ITmedia reports ChatGPT's and Claude's Microsoft 365 Connector A ready-made link that lets an AI assistant use another app or service, like Grab or Gmail, with your permission. stopped working, with Anthropic saying most mail, calendar and search requests failed. Ordinary chats outside those connectors were not part of the incident.
Traffic returned after about 18 hours of downtime, internet monitoring firm Kentik's Doug Madory said, according to TechCrunch. Microsoft's Copilot chief Charles Lamanna said teams are reviewing how to prevent a repeat.
Why it matters: AI assistants built into office tools are only as reliable as the cloud under them. Keep offline copies of key contacts and files, and agree a backup channel such as phone or WhatsApp for days when Outlook and Teams are down.
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
Robots
74.8%
Astra agent score on a simulated robot benchmark (Nvidia study)
Robots
OpenAI breaks up robotics team, betting its main models can run robots
OpenAI has disbanded its stand-alone robotics team and moved the staff into other groups, Fortune reported on 13 November (US time), and OpenAI confirmed the change. The company says robotics is still a core goal and the move is meant to speed it up.
Key learnings
Robotics researchers now sit in OpenAI's main research group under chief research officer Mark Chen. Applied robotics and hardware staff moved to the applied group under president Greg Brockman.
OpenAI's reason: its general-purpose models, including GPT-6 Astra, can now be applied directly to robots. Fortune cites an Nvidia study in which an Astra-powered AI agent An AI program that can take actions for you, like writing code, filling forms or moving items, not just answering questions. scored 74.8% on a simulated robot Benchmark A standard test used to compare AI models, like an exam everyone takes so scores can be compared., against 51.0% for the best specialised robot models. That is Simulation Testing a robot or system in a virtual world on a computer before trying it in real life., not proof in real factories.
OpenAI is still building its own Humanoid robot A robot built roughly in the shape of a person, with arms, a torso and often legs or wheels., with CEO Sam Altman targeting a first public demo in 2027. The change came so recently that OpenAI's booth at the CoRL robotics conference in Austin this week stood empty.
Not everyone agrees: Carnegie Mellon's Guanya Shi notes large models need a constant Cloud computing Renting computers in someone else's data centre over the internet instead of buying your own. connection, so small models running on the robot still matter, and MIT's Russ Tedrake says robot makers' real-world data remains an advantage.
Why it matters: If big general models can control robots, factories and warehouses may not need a separate robot brain for every machine, which could cut costs. Treat the benchmark lead as early: it was measured in simulation, and robots on the floor still need control that works offline.
Exclusive: OpenAI disbands its stand-alone robotics team as its general models start running robots · FortuneRead ↗
Was this useful?Tap to vote. Hold, right-click or press Shift+Enter to add a reason or a note first.
Developer tools
1,000
questions allowed in one request
Merged from 2 sources
Developer tools
Google's Gemini Decision 1 answers yes-or-no questions without writing text
Google launched the Gemini Decisions API Application programming interface: a door that lets one piece of software ask another for data or actions. on 11 November (US time), powered by a new model, Gemini Decision 1. Instead of writing an answer word by word, it picks from options you give it and returns probabilities plus a Confidence score A number, often from 0 to 1, showing how sure a model is about its answer; it does not prove the answer is right. from 0 to 1.
Key learnings
Three question types: yes/no, multiple choice (2 to 255 options) and a score on a scale you define. One request can ask up to 1,000 questions about a single input of text, images, audio, video or PDF.
Google's figures: US$0.10 per million input 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., nothing for output, and 150 to 250 milliseconds per request. Google says it is nearly 5x cheaper than Gemini 3.5 Flash-Lite on its Benchmark A standard test used to compare AI models, like an exam everyone takes so scores can be compared. suite; these are Google's own claims.
In a hands-on video on 13 November, the Prompt Engineering channel asked about the weather in Paris and the model picked 'technical support' with 92% confidence, because no option fitted. Adding an 'other' option fixed it: confidence only tells you which of your options is closest.
The model cannot explain its answers because it does not write text. The tester's rules: give every choice a way out, send only what each decision needs, ask small questions and test on your own data.
Why it matters: Many business tasks are small decisions, such as which team gets a Trouble ticket A logged request about a problem, so a support team can track it until it is fixed., and a cheap, fast Classifier A model that sorts an input into one of a fixed set of labels, such as spam or not spam. like this can handle them and pass unsure cases to a person. Set a confidence cut-off and keep a normal chatbot or a human for anything that needs a reason.
AI agents
