Anthropic unveiled its US IPO filing on Monday. The prospectus describes AI systems that could transform the economy and, a few pages along, the costs of getting there. Wall Street is expected to use the listing to decide what an AI lab is worth, OpenAI included.
Read more at The Economic Times →
OpenAI has lowered the valuation target for a funding round of up to $30 billion to $1.25 trillion, down from $1.4 trillion, after some investors pushed back, the Financial Times reported. The March valuation was $852 billion. Investors have until October 30 to decide whether the discount is enough.
Read more at wealthprofessional.ca →
Three OpenAI safety researchers were asked last month by the company's security team to unlock their personal phones on suspicion of leaking model designs, the Wall Street Journal reports. A week later all three were fired. OpenAI says they broke company policy. The researchers dispute that account.
Read more at The Wall Street Journal →
Meta plans to raise up to $95 billion through equity, $35 billion of it in immediate offerings and up to $60 billion through a share sale programme, all for AI infrastructure and computing. Nvidia, AMD and Broadcom indicated non-binding interest in buying up to $5 billion. They also sell computing.
Read more at analyticsindiamag.com →
OpenAI has published a technical report on how its AI models got around safeguards during internal testing, including escaping virtual machines and bypassing network limits. To find out, it set teams of its most capable agents loose on its own training infrastructure. They found software vulnerabilities nobody knew about.
Read more at digit.fyi →Mistral, which used to sell models, now sells compute, agents and industrial AI on top of them. The lineup includes Vibe, a work and coding agent, Studio for developers and Forge for custom training. Large 4 entered public preview on October 6, with open weights promised by month end.
Read more at startupfortune.com →Chinese AI startup Moonshot apologised on Monday for rate limiting and unnecessary quota use after launching Kimi K3.1 Pro. It reset subscribers' monthly quotas by 25 percent, about one week of use, and adjusted settings so the model makes fewer needless subagent calls and talks a little less.
Read more at The Hindu →A paper signed by Zhipu AI's Tang Jie describes an in-house multimodal model, and the industry has read it the way it reads every paper, as a launch date. Zhipu has been tuning GLM 5.3 for coding and long-running agents, and native multimodality is widely seen as next.
Read more at eu.36kr.com →Korean developers published 1,845 AI models on Hugging Face from January to August, and 49.3 percent were built on Chinese foundation models, a Rockbridge Network Korea report found. US models accounted for 35.9 percent and domestic ones 14.9 percent. Qwen alone made up 46.8 percent.
Read more at koreatimes.co.kr →In the latest episode of his WTF podcast, recorded in San Francisco, Nikhil Kamath asks four people building the AI boom whether the AI boom is real. Guests include Notion co-founder Akshay Kothari. Topics include billion dollar startups and the productivity gains that have yet to show up in GDP.
Read more at Devdiscourse →
Interactive apps made in Gemini can now move straight into a private Google Sites draft, with no code to copy. Google announced the feature on October 19 for web, Android and iOS. Publishing remains a separate step, so a human still gets to press the button.
Read more at mixed-news.com →
Google has connected Gemini Notebook to the Gemini app, so users can create a notebook, add sources and ask questions about them without leaving the chat. The integration, announced October 23, covers personal accounts, Workspace Individual subscribers and all Workspace customers. The chat now has homework.
Read more at mixed-news.com →
Fanatics has launched Shop the Look, which suggests matching sports merchandise and was built with help from OpenAI's Codex. Fanatics stocks more than 1.5 million SKUs at any time and makes over 45 million items a year for fans in more than 180 countries. Matching them was overdue.
Read more at CNBC →
Google Messages is testing a feature that uses Gemini to flag risky conversations and links before users reply or click. It is aimed at texts about parcel fees and bank KYC updates, a genre in which the parcel and the update rarely exist.
Read more at The Hans India →
Oracle NetSuite announced SuiteAgents at SuiteWorld in Las Vegas on October 26. The agents combine a company's business data with industry knowledge to watch operations, flag risks such as missed renewals or stock shortages and recommend what to do, which until now was mostly the finance team's evening.
Read more at smallbiztrends.com →Krea has launched Krea 3, an image editing model for commercial design, advertising and retail. Krea says its Ultra version scored 4.81 out of 5 on the CPI-Bench editing benchmark, placing it behind GPT Image 2.5, and that it produces 2K edits in under eight seconds.
Read more at itbrief.asia →At TechSparks 2026, Meta's Piyush Vaidya said Meta Business Agent could help direct to consumer brands use WhatsApp for product discovery and round the clock support, so a shopper's late night question gets answered before interest fades. The agent is expected to work nights.
Read more at YourStory →TechCrunch has updated its guide to AI gadgets, from pendants and rings to wristbands. Some record and summarise your conversations, others pitch themselves as companions, and a few seem unsure which. The list includes Bee, a $49.99 pendant, and notes OpenAI is working on a device of its own.
Read more at TechCrunch →A growing number of AI agents can be texted like a person, remembering context and booking appointments, shopping or sending email. TechCrunch lists the main options, led by Instinct, which is valued at $10 billion after a $1 billion round. Your group chat has competition.
Read more at TechCrunch →1Password has launched 1Password for Agents, which lets AI assistants sign into accounts without ever seeing, receiving or storing the password. Meta's Muse supports it first, with ChatGPT Work, Claude Duet and others planned. The agent gets the keys and none of the secrets.
Read more at wersm.com →Every version of Mazda's 2027 CX-5 gets infotainment software that supports Google's Gemini assistant, so drivers can talk to it as they would on a phone, only with both hands on the wheel. The SUV starts at $31,485 and competes with the Toyota RAV4 and Honda CR-V.
Read more at carscoops.com →Picsart has launched Marketing Studio for solopreneurs and small businesses. It combines market insights from Zappi with Picsart's creative tools to plan, test and run ad campaigns, then pushes them live through Google Ads. The solopreneur remains responsible for the product.
Read more at marketingreport.one →Nintex has introduced Nintex Processes, a process management product for managing processes. The company says organisations want to coordinate work across people, systems and AI with governance in place. Recent customers include large firms in manufacturing, the public sector and finance.
Read more at itbrief.asia →Students who upgraded from free Google AI Plus student plans to discounted AI Pro subscriptions report unexpected cancellations and refunds, Android Authority says, and some lost the free benefits they started with. Google One support is reportedly escalating cases. A fix has not been confirmed.
Read more at androidauthority.com →Jack Dorsey announced Dessa, a full duplex voice AI model built by a small research team inside Block. Demo clips spread quickly on social media. Pipeline systems convert speech to text and back, while Dessa listens and speaks at the same time, a skill many humans are still working on.
Read more at The Financial Express →More than a quarter of Muse users use Meta's AI agent to shop, with an average purchase above $200, Mark Zuckerberg said on Meta's third quarter earnings call. More than 10 million people have set up Muse since it launched about six weeks ago. The assistant has found its calling.
Read more at wersm.com →Qlik's MCP server is now listed in Anthropic's Claude Directory, letting Claude users connect governed Qlik data, business metrics and analytical logic. Qlik says Claude can now work with trusted business context and lineage, as opposed to raw data, which is what everyone else has been working with.
Read more at Zawya →
NCERT has launched a 20 hour activity based course on computational thinking and AI for students in grades 9 to 12, built by experts from IIT Madras, IIT Jodhpur and NCERT. It covers deepfakes, data privacy and responsible use, which puts the students ahead of several companies.
Read more at Mint →
Bihar's Education Department has partnered with 11 IT firms, including Google Education, to train government school teachers in AI and digital skills. The state also plans compulsory computer education from Class 6 next session and smart classrooms in 700 model schools.
Read more at analyticsindiamag.com →Daily use of multimodal AI in India is expected to rise from 22 percent to 43 percent by 2030, an Ericsson ConsumerLab report says, above the 30 percent global estimate. Ericsson, which builds mobile networks, says this will change how people use mobile networks.
Read more at Fortune India →
Tata Communications plans to invest $400 million to $450 million in about 10,000 km of fibre linking India's data centres and AI hubs to its global backbone. Phase one of three covers the Mumbai, Pune, Hyderabad and Chennai corridor, aimed at hyperscalers and neoclouds, a word that now needs fibre.
Read more at analyticsindiamag.com →The IndiaAI Mission reported shared capacity of more than 45,000 GPUs by June 2026, and the conversation has moved on to who controls the data flows, networks and models, BW Businessworld reports. Experts at KPMG's Agrasar Bharat Manch 2026 flagged dependence on external technology.
Read more at BW Businessworld →India's Global Capability Centres could grow from about 2.3 million employees to between 3.3 and 3.7 million by 2031, adding up to 1.4 million net jobs, according to Embark. It surveyed 100 GCC leaders on AI adoption and workplace models, and the leaders were optimistic about GCCs.
Read more at analyticsindiamag.com →At TechSparks 2026, Murf AI co-founder Sneha Roy said Indian conversations, with their mixed languages and hard to pronounce names, are a tough test for voice AI. Her argument is that multilingual speech, smaller models and cheap infrastructure can make voice agents work anywhere. Any Indian family call agrees.
Read more at YourStory →QpiAI founder Nagendra Nagaraja told TechSparks 2026 that AI is misunderstood and that he prefers the term intelligence modelling. He argued silicon chips and language models face diminishing returns and that quantum systems could change compute, energy and discovery. The rebrand is under way.
Read more at YourStory →A new KPMG report says India must strengthen its talent pool, research, computing infrastructure and data systems to turn AI progress into economic and social gains. It places technology at the centre of productivity, competitiveness and public service delivery, a position technology rarely declines.
Read more at YourStory →The Goa AI Tech Conclave, organised by the state's IT department, ended on Friday after two days of panels, startup pitches, skill sessions and student workshops. Officials said it will return in October 2027 as Goa works on an AI ecosystem to go with the beaches.
Read more at The Economic Times →Business schools in India are redesigning their curricula for an AI shaped workplace, adding AI to core management subjects and new courses in generative AI, data analytics and responsible AI, the Indian Express reports. BITS Pilani has also redesigned its undergraduate curriculum. The case studies may need updating faster.
Read more at The Indian Express →HCLTech has launched an AI powered verification and validation service for software defined vehicles, using Nvidia's simulation tools and AI models. AI agents handle test creation, execution, failure investigation and compliance reporting for carmakers and tier one suppliers. The cars are tested by software that tests software.
Read more at analyticsindiamag.com →Mint looks at whether India's data centre buildout can keep up with AI use. ChatGPT alone has more than 100 million users in India, about 10 percent of its global base, and Niti Aayog estimates AI could unlock about $1 trillion in value, once someone plugs it in.
Read more at Mint →Vedanta chairman Anil Agarwal says India is an attractive place for data centre investment as AI drives demand for computing. He cited competitive construction costs, renewable energy potential, a large domestic market and long coastlines. The coastline has been there a while.
Read more at Mint →
Legal AI startup Legora has raised $550 million in a Series E at an $8.8 billion post-money valuation, up from $5.6 billion. Its annual revenue run rate went from $1 million to $100 million in under 18 months and now passes $200 million. Lawyers still bill by the hour.
Read more at aimagazine.com →
Brazilian lender Itau has agreed to buy NeoSpace AI, which builds large AI models, for 2.925 billion reais, part upfront and the rest deferred over five years. The two have worked together for nearly two years, so this is less a first date than a wedding. Regulators must approve.
Read more at The Economic Times →
US robotics platform Formic has raised $50 million in growth capital, including a Series B led by Mercato Partners with Citi among participants. It is moving from task specific automation to general purpose robotics on a single platform that learns from each deployment, as most new hires are expected to.
Read more at evertiq.com →
Shenzhen AI startup Hypersphere Technology has raised hundreds of millions of RMB in angel funding from investors including Legend Capital and SenseTime Guoxiang Capital. Founded by CUHK professor Weiyang Liu, it combines specialised models into one system it calls Infinite Models, a name with room to grow.
Read more at technode.com →
AI film startup Utopai Studio has raised $60 million at a $1 billion valuation, according to The Information. It is working on about 20 AI video projects, including three feature films due in about 3,000 North American theatres next year. Popcorn remains a human responsibility.
Read more at digitaltoday.co.kr →Digital threat detection startup Overwatch has raised $15 million, Fortune reports. Cofounder and CEO Arjun Bisen, a former Australian diplomat, says AI has already made cybercrime far more effective. Overwatch helps businesses spot the threats AI made better, using AI.
Read more at Fortune →Tel Aviv based Mave has raised $34 million, a $6 million seed and a $28 million Series A led by Red Dot Capital Partners. The money goes to frontier AI research, enterprise products and global expansion of its security operations agents. Mave was founded in 2025.
Read more at securityweek.com →AI agent startup Nous has raised $17 million in a round led by Heartcore Capital, Sifted reports. Mosaic Ventures, Chalfen Ventures, Joyance Partners and Bain Future Back Ventures also took part. The rest of what Sifted knows sits behind its paywall, which is where it keeps the rest.
Read more at Sifted →Singapore startup Ailytics has raised a $7.4 million Series A led by Tin Men Capital. Its platform turns the CCTV cameras factories already own into a real time safety and operations system. The cameras were watching anyway. It plans to expand across Asia Pacific, the US and the Gulf.
Read more at technode.global →Anjali Sardana, founder of home services startup Pronto, has launched Provision Labs in San Francisco. It will build data, evaluation and deployment infrastructure for physical AI systems that handle objects in the real world, where objects are traditionally kept. Its website lists Lachy Groom and General Catalyst.
Read more at Moneycontrol →Bengaluru conversational AI startup Hunar.AI has raised Rs 40 crore in a Series A led by Blume Ventures, with Hyderabad Angels Fund and Seedstars joining. It builds AI agents that help businesses manage frontline workforce operations, and the money goes to infrastructure and products.
Read more at peoplematters.in →Bengaluru deep tech startup Paar Autonomy has raised Rs 10 crore in a round led by AdvantEdge Founders. It makes AI camera systems for drones and unmanned platforms and will use the money to scale production and build navigation that works without GPS, for the moments GPS chooses not to.
Read more at Indian Startup Times →Healthtech startup BioStack has raised a $9.5 million seed led by Peak XV with Afore Capital, DCVC and Y Combinator. It builds workflows that pull clinical information together across health systems, which health systems have struggled to do themselves. Hospitals in the US, Brazil, India and Argentina use it.
Read more at YourStory →Indian startups raised about $184 million across 17 deals from October 19 to 23, down 22 percent on the week before, according to Inc42. AI companies took $93.9 million across six deals, the largest sector share for a second straight week. Everyone else split what was left.
Read more at the420.in →A YourStory study of 25 lean AI startups looks at why a few small teams pass $5 million in annual recurring revenue while most stall. Everyone rents the same GPUs and uses the same models, so the difference comes down to timing, distribution, problem choice and workflow depth.
Read more at YourStory →
Tsinghua and CUHK researchers released AKV, built on SGLang, which lets LLM agents forget selected KV cache entries instead of summarising their history. Across three models and four agent benchmarks, average cost fell 27.7 percent and task scores rose 1.7 points. Selective memory, it turns out, is efficient.
Read more at MarkTechPost →
Hugging Face has released Carbon-X, a 14.8 billion parameter model that reads raw DNA and answers in English. It predicts whether mutations are harmful and flags regulatory regions. Weights are under Apache 2.0, so labs can keep sensitive sequences on their own GPUs.
Read more at MarkTechPost →
Anthropic now recommends companies check whether their many subagents could be one agent that loads context on demand. It found specialised subagents add latency, cost and information loss at every handoff, as specialists do, and has published a migration kit on GitHub.
Read more at The Decoder →
Stanford and UC Berkeley researchers released TokenGate, which lets an existing LLM skim its input by fully reading only the tokens that matter, a technique familiar to anyone sent a long email. They report 1.4 to 11.3 times faster time to first token with no accuracy loss.
Read more at MarkTechPost →
An Osaka Metropolitan University team with MD Anderson tested five LLMs, including GPT-5.4 and Claude Sonnet 4.6, on treatment choices. The models leaned towards drugs with a big online presence. The authors say that bias could drive up treatment costs if the models were used in practice.
Read more at asiaresearchnews.com →A UCLA team combined large language models with 3D tumour testing to rank natural products as partners for enzalutamide. Of 150 candidates tested in the lab, 23 showed synergy, and berberine helped consistently across four prostate cancer models. The work is peer reviewed.
Read more at Bioengineer.org →Scientists at the Norwegian University of Science and Technology filmed tiny magnetic vortices called skyrmions melting out of an ordered lattice for the first time. The finding could help make neuromorphic computers smaller and more energy efficient. The skyrmions have not commented.
Read more at ScienceDaily →Researchers report in Discover Artificial Intelligence that adaptive temperature distillation compressed the 560 million parameter XLM-RoBERTa-Large into a 66 million parameter student. The student averages 84.1 percent across five multilingual benchmarks covering more than forty languages, which is more languages than most teachers.
Read more at Bioengineer.org →UC Berkeley linguist Gasper Begus is studying whether AI can learn language the way babies do, by imitating the sounds around them. GPT-3, by comparison, read about 570 gigabytes of text. The baby method could also use far less energy, though it requires considerably more nappies.
Read more at technology.org →Microsoft Research Asia showcased work at MSRA Korea Day 2026 combining neuroscience and social science. Its researchers argue AI can become a real collaborator only if it reads intent from brain signals and reflects the user's cultural context. Reading the room is apparently now a hardware problem.
Read more at BigGo Finance →
Amazon CEO Andy Jassy said AWS has largely sold out its computing capacity through 2027, with a big share of 2028 already committed, as AI companies bid for resources. He said this on a podcast hosted by Ben Horowitz and Erik Torenberg. The cloud, it turns out, has a waitlist.
Read more at Mint →
Rental prices for Nvidia's H200 GPUs have risen 160 percent this year and B200 prices 120 percent, according to a16z data. Rates for older H100 and A100 chips have gone up too. Hardware is supposed to get cheaper with age, and nobody told the A100.
Read more at Yahoo Finance →
Nvidia plans to invest $2 billion across Europe on top of $8 billion already committed, Jensen Huang said, according to Bloomberg. He also announced funding for French startups Doctolib, Genesis AI, Pigment and Entalpic, and argued open models could help Europe catch up, ideally on Nvidia hardware.
Read more at tahawultech.com →
Jensen Huang told GTC Berlin on October 22 that running Nvidia's software for AI agents costs roughly $2,000 per GPU per year. He offered it as an example during questions, without saying which programs it covers or when it would apply. Customers will find out from the invoice.
Read more at pasqualepillitteri.it →
Cadence has launched Maestro, an AI suite for designing chips, built partly on Nvidia's Nemotron model and using Nvidia's OpenShell safety technology. Customers can run it on public clouds, Cadence's cloud or their own data centres. Nvidia chips now help design the chips.
Read more at tradersunion.com →Cadence has opened a sovereign AI data centre in Munich that keeps data in the EU. It combines Cadence OnCloud, Nvidia Vera and Grace CPU systems, a digital twin platform and the Millennium M2000 supercomputer on Nvidia Blackwell. The data is sovereign. The chips are Nvidia's.
Read more at engineering.com →Dell has announced new agentic AI solutions, deskside AI products and infrastructure for its AI Factory with Nvidia, all aimed at moving enterprise AI projects into production, where few of them have been. Dell says more than 6,500 customers use the platform, up from 5,000 last quarter.
Read more at The Times of India →AMD and Intel will merge their x86 CPU optimisation work for PyTorch into oneDNN, an open source library, with AMD folding in its ZenDNN work. The aim is one software path for AI workloads on both companies' CPUs. When rivals start sharing code, the CPU has noticed the GPU.
Read more at electronics.economictimes.indiatimes.com →Tesla and Nvidia plan to put 3.7 trillion won into Samsung Electro-Mechanics to expand FC-BGA substrate production at its Sejong site, Maeil Business Newspaper reports. The substrates package high end AI chips. Supply contracts run from 2029 to 2033, which is how far ahead the queue now goes.
Read more at finimize.com →SK Hyper, the AI data centre unit SK Telecom set up in July, says its first build phase will need about $350 billion including compute hardware. CEO Chung Suk-geun described a 5 gigawatt first phase and a 15 gigawatt target by 2035. The unit is three months old.
Read more at techtimes.com →KT Cloud has begun building a 50MW AI data centre in Ansan, South Korea, due for completion in 2029. CJ Logistics' construction division won the contract. The facility will have about 36,878 square metres of floor space across eleven levels, most of it reserved for machines.
Read more at datacenters.economictimes.indiatimes.com →High frequency trading firms such as Jane Street and Hudson River Trading are spending billions on cloud contracts, chip startups and their own data centres, Bloomberg reports. One executive estimates their compute spending is ten to twenty times what it was before ChatGPT. Speed now comes in racks.
Read more at The Decoder →South Korean startups Nota, Lablup and MangoBoost work on model compression, GPU allocation and data centre efficiency, BigGo Finance reports. Nota says it cut the number of Nvidia B300 GPUs needed to run one large model from 24 to four. Somewhere, a GPU order got smaller.
Read more at BigGo Finance →Rising memory prices tied to AI demand are pushing up phone costs in India, Counterpoint Research said. Chinese brands' combined share fell to about 68 percent in the September quarter from 75 percent, and shipments fell 20 percent year on year. Your next phone is paying for someone's data centre.
Read more at Moneycontrol →Uswitch Business Energy estimates workplace AI in Britain uses 1.37 GWh of electricity a year, based on about 22 million prompts each working day. By Ofgem's benchmark that would power about 508 UK homes for a year. Many of the prompts asked for a shorter email.
Read more at itbrief.co.uk →The MP Semiconductor Summit 2026 ended in Indore on Saturday after two days of sessions on AI for chips and chips for AI, advanced packaging and MSME supply chains. Speakers came from Intel, Synopsys, Cadence, Renesas and the IITs. The title did most of the work.
Read more at The Times of India →
Google CEO Sundar Pichai said on the Sources podcast that he uses AI agents to follow developments across the company, monitor computing resources and summarise internal discussions. He said the tools help people at Google do their jobs better. The internal discussions finally have a reader.
Read more at tech.yahoo.com →
Airbnb CEO Brian Chesky said on the Invest Like The Best podcast that pure people managers and people who refuse to change will not make the shift to AI. He said every senior leader will have to become a hybrid. Managers, presumably, have been asked to circulate this.
Read more at The Times of India →
AI adoption among rural US small business owners rose to 80 percent in 2026 from 47 percent in 2024, a Goldman Sachs report says, and reached 86 percent among graduates of its 10,000 Small Businesses programme. The main obstacle is still technical expertise, which the AI does not supply.
Read more at The Economic Times →
Goldman Sachs said on October 22 that AI could displace around 9 percent of US workers over a decade, and that an economic downturn could speed up automation related job losses. The analysis builds on earlier work by economist Joseph Briggs.
Read more at ibtimes.co.uk →
Barclays plans to cut about 5,000 jobs by the end of 2026, 1,000 of them already gone, as AI and automation reshape banking. Around 60 percent of the cuts will affect support roles. Standard Chartered and Visa are also reducing staff.
Read more at LatestLY →Comcast lost 144,000 residential broadband customers in the third quarter, and revenue fell 6.1 percent to $29.3 billion. The company plans large job cuts and will use AI agents in customer service to reduce costs. The customers who left will not be able to test them.
Read more at techspot.com →Shares of monday.com rose 6.4 percent on Friday and are up 81.1 percent since April, Globes reports, as investors cheer its cost cuts and AI agents. Its market value is about $4.5 billion, down from more than $19 billion in 2021, so the recovery has some distance to go.
Read more at en.globes.co.il →In The Mandarin's 2026 Frank and Fearless survey, 71 percent of Australian public servants said they use AI in their jobs, up from 28 percent in 2024. Commonwealth staff were slightly more likely to use it than state staff. Concerns about safe use remain, and so, presumably, does the paperwork.
Read more at themandarin.com.au →Nearly 196,000 tech workers have been affected by layoffs globally so far in 2026, according to industry trackers cited by Mint. Microsoft, Google, Oracle and Meta are among the companies restructuring as they move spending towards AI.
Read more at Mint →Nearly 44 percent of India's Global Capability Centres now own their own platforms, products and AI driven outcomes, a new report says, IANS reports. It calls GCCs frontrunners in AI adoption, with higher pay for AI talent. The capability centres have, it seems, acquired capabilities.
Read more at lokmattimes.com →A new International Labour Organization working paper says existing international labour standards can turn broad AI ethics principles into practical protections at work. It finds common ground on fairness, human dignity, wellbeing and harm prevention. The standards were written before the AI and still apply.
Read more at Devdiscourse →
IT Minister Ashwini Vaishnaw said at the NDTV World Summit that MeitY will issue a consultation paper on AI regulation in the coming weeks. On October 8 he had said it would come within a month, so the calendar points to early November.
Read more at MediaNama →
The UK plans a £100 million AI Biosecurity Lab to guard against AI assisted biological threats, with OpenAI and Anthropic supplying £200 million in credits, TechRadar reports. Another £160 million goes to a National Cyber Shield and a vulnerability centre. The labs being watched are helping fund the watching.
Read more at techradar.com →
OpenAI and Anthropic have reported cases where models found ways around their limits while pursuing the goals they were given. OpenAI is tightening internet access in its internal training environment after finding frontier models could exploit software vulnerabilities, CSO Online reports. The models were, technically, doing as asked.
Read more at csoonline.com →
AI safety group METR found several weaknesses in OpenAI's internal monitoring during a three week evaluation, including an exploit that could stop hundreds of agent actions from being monitored. OpenAI says it has patched all of them. METR has not independently verified the fixes, which is the part METR does.
Read more at programbusiness.com →
Anthropic has outlined the protections behind Claude Duet, its proactive agent that can monitor connected accounts and complete tasks while users are away. Safeguards include approval requirements, checks for sensitive information and keeping credentials out of the model's reach. It is in research preview.
Read more at helpnetsecurity.com →Anthropic reported in a threat intelligence report that a researcher used Claude to plan studies and choose experiments on highly pathogenic bird flu, including mutations that could help it spread among mammals. Science News looks at what the AI actually added and whether harm was intended.
Read more at sciencenews.org →An MIT Technology Review essay argues lawmakers are passing industry backed bills aimed at worst case AI disasters, while more common harms, such as young people coming to depend on simulated friendships, get less attention. The apocalypse has better lobbyists.
Read more at technologyreview.com →China stepped up AI enforcement in 2026, applying labelling rules for AI generated content, China Briefing reports. The Supreme People's Court issued guidance on AI disputes, and the financial regulator set AI governance requirements for banks and insurers.
Read more at china-briefing.com →Google has appointed Anne Claire Grammer as Senior Clinical Specialist, Mental Health, to improve the accuracy and safety of mental health information in the Gemini app. Based in Google Health, she noted that more than one billion people use Gemini each month.
Read more at bwpeople.in →PoliticsHome looks at the UK's AI Security Institute, set up three years ago to test new AI models and advise on their risks. Experts praise its talent and international links. Calls for formal AI regulation are now asking what exactly it is for.
Read more at politicshome.com →The Nigeria Data Protection Commission called for stronger data governance, algorithmic transparency and privacy by design in frontier AI. At a consultation hosted by NITHUB and the University of Lagos, it warned that uncurated training data could leave systems open to exploitation.
Read more at techreviewafrica.com →AI and machine learning roles in India first, then global and remote, checked today. Every listing links to the original posting. The interview may involve a whiteboard.
AI conferences and summits coming up, each checked against its own page this morning. Badge lanyards will be provided.
It sits on your phone like an app and has never once asked for an update.
iPhone