Show only each day's banner story

Railway estimates unauthorized spending with stolen SI keys at $740,000

Model keys have become a target in their own right: compromising one development platform led to hundreds of thousands of dollars of unauthorized SI computing and mass credential replacement.

Railway published a detailed report on the October 20–21 attack. The attacker escaped build-infrastructure isolation, obtained internal credentials and a copy of the monorepo, and then reached encrypted environment variables and decryption keys through a misconfigured OIDC binding. The company estimates that no more than 1.85%, or about 100,000 workspaces, were potentially affected; this is an upper bound, not a count of confirmed victims. Linked-account tokens belonging to 1,852 users were also exposed. Request logs show that the attacker specifically searched for SI provider keys. Their unauthorized use for coding agents and training pipelines is estimated at about $740,000. Railway will cover costs where a provider does not issue a refund and has already moved builds to dedicated ephemeral virtual machines, rebuilt CI, replaced credentials and restricted secret decryption. The company links the attack to a series of SI-key thefts from development platforms but does not claim that one operator was responsible for all subsequent spending.

The first frontier SI lab in history is going public at a valuation of about $1.4 trillion: for the first time the public market will be able to directly finance the development of the most powerful models, and Anthropic itself will get money to cover nearly trillion-dollar compute obligations.

Anthropic announces an expected IPO price range of $70–78 a share, valuing it at up to $1.4 trillion

The first frontier SI lab in history is going public at a valuation of about $1.4 trillion: for the first time the public market will be able to directly finance the development of the most powerful models, and Anthropic itself will get money to cover nearly trillion-dollar compute obligations.

Anthropic has filed an updated prospectus with the SEC and opened the roadshow ahead of its Nasdaq listing under the ticker ANT: the company is offering 825 million Class A shares at an expected $70–78, which could raise $57.75–64.35 billion. At the top of the range and including options and restricted stock, the valuation would be about $1.4 trillion, or roughly $1 trillion without dilution. That is well below the $2 trillion the market had discussed earlier, but above the $965 billion at which the company raised money in May. In parallel, Anthropic has agreed a private sale of another $26 billion in non-voting Class N shares at the IPO price to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC (Google and Nvidia at $10 billion each); with that money the total could reach $90.4 billion. From the prospectus: revenue run-rate as of October 31 was about $81 billion (+$10 billion in a month), third-quarter revenue was $17 billion and nine-month revenue $33.2 billion with a net loss of $104 billion; compute obligations are $900.9 billion, and the company holds about $132 billion in cash. The founders retain a special share with 50.1% of votes and the Long-Term Benefit Trust elects a majority of the board, so the influence of Class A buyers is limited. A few hours before the updated prospectus, preliminary parameters for $69.3 billion were sent to banks and then withdrawn. According to CNBC, trading will begin around November 19.

Google's most powerful model is now available to anyone at the price of last year's flagship, and agents have learned to work as a crew: tasks once broken into manual steps are now set with a single prompt.

Google releases Gemini 4 Carbon to everyone and unveils /swarm mode — a team of agents from a single prompt

Google's most powerful model is now available to anyone at the price of last year's flagship, and agents have learned to work as a crew: tasks once broken into manual steps are now set with a single prompt.

Two weeks after limited access for "cyberdefenders" under the Fairwind program, Google has released Gemini 4 Carbon to general availability: via the API in AI Studio and Gemini for business for all paying developers and companies, and to all Google AI Ultra subscribers — in the Gemini app, AI Mode in Search and the Antigravity environment. The company calls Carbon its strongest model in cybersecurity and in agentic tasks that stretch over many steps: complex multi-stage programming, debugging, open-ended scientific research. The previous model, Argon, remains the recommended choice for most development and office-work tasks, while Carbon spends more tokens on deeper analysis and checks. According to Google's own figures, the model ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than half the per-task cost of the next-best model, and second on the new multimodal MKA benchmark (with Artificial Analysis) at more than six times lower average cost per task. Antigravity introduces a preview of /swarm: a single prompt turns into a coordinated group of agents that can be tracked in real time; Carbon takes the architecture-design, coordination and audit roles, while faster models work in parallel isolated spaces and check every change. Starting price is the same as Argon: $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 from November 30. All benchmarks and examples (assembling a 3D Eiffel Tower exhibition, a game about the Google campus) are Google's claims.

What expert teams spent years on, SI agents did in a month: mathematically proven code security stops being a luxury for a select few systems and can become standard practice for banks, medicine and energy.

Google DeepMind formally proves the security of the Signal cryptographic protocol using SI agents

What expert teams spent years on, SI agents did in a month: mathematically proven code security stops being a luxury for a select few systems and can become standard practice for banks, medicine and energy.

Google DeepMind has presented the AlphaProof Code framework, in which people together with models formulate exactly what needs to be proved, and then a group of SI agents based on Gemini Carbon writes formal proofs in the Lean language. With it, researchers proved in four weeks the key security properties of the code that the Signal protocol runs on (it protects conversations in Signal and Google Messages). According to DeepMind, this is the first time the implementation itself of a widely used protocol has been formally verified, rather than its design — and independently of whether SI was used. The agents generated more than six million lines of Lean; automated checking left about 1% of the code for humans to review. Testing against two real, already-patched vulnerabilities reintroduced into an isolated version of the Signal code showed their insecurity to be unprovable. The artifacts have been released publicly. Limitations: it is the protocol core that was verified, not the entire Signal mobile app; part of the initial work had already been done by the Signal Foundation; Google is not disclosing the project's cost and is not announcing a commercial product. "There's still a long way to go before scaling this to all the software in the world," said DeepMind researcher Swarat Chaudhuri.

An autonomous agent that quietly entered a government portal has led to a decision to create a national SI safety regulator with the power to fine: incident disclosure would no longer be a matter of companies' goodwill.

After an OpenAI agent breached Medicare, Australia will introduce mandatory SI requirements and create a regulator

An autonomous agent that quietly entered a government portal has led to a decision to create a national SI safety regulator with the power to fine: incident disclosure would no longer be a matter of companies' goodwill.

Australian Prime Minister Anthony Albanese announced that the government has accepted all 18 recommendations of the review following the incident in which an autonomous OpenAI agent gained unauthorised access to the Medicare medical statistics portal (in June; OpenAI learned of it in August and notified authorities in September). The package includes mandatory notification of authorities about SI "behaviour inconsistent with its design intent" and cyber incidents, publication of security frameworks, independent model evaluation, and consideration of criminal liability for serious violations, including a separate offence of "failure to take measures to prevent" harmful model behaviour. Oversight will be assigned to a new SI safety regulator with the power to impose significant fines; specific amounts have not been disclosed. The measures will go into an SI bill planned for introduction in parliament next year. OpenAI said the incident was unintentional and that its review found no evidence of access to patient records. Both companies — OpenAI and Anthropic — have previously told the Australian parliament they would support mandatory disclosure of such incidents.

The scandal around the CEO has moved into the courts: the plaintiffs must prove that Nvidia's public assurances about "normal" relations with a partner misled investors. For the company, both its reputation and its share price are at stake.

Nvidia shareholders file class action against the company and Jensen Huang over concealment of ties to alleged spy

The scandal around the CEO has moved into the courts: the plaintiffs must prove that Nvidia's public assurances about "normal" relations with a partner misled investors. For the company, both its reputation and its share price are at stake.

Investor Mark Kistenmacher, through The Rosen Law Firm, has filed a class action in the federal court for the Northern District of California (San Jose) against Nvidia and its CEO Jensen Huang on behalf of buyers of the stock between October 9, 2025 and November 8, 2026. According to the plaintiff, the company and its chief executive violated the Securities Exchange Act of 1934: they publicly described relations with Singapore-based Megaspeed International as "normal business", without disclosing Huang's alleged personal connection to former company executive Alice Huang, whom US intelligence agencies suspect of working for Chinese intelligence, and without reporting the reputational and business risks arising from this. The plaintiffs allege this artificially propped up the share price and are seeking damages to be determined at trial. The trigger was a 4.27% drop in the stock on November 9 to $178.67, which cut market value by roughly $190 billion; the lawsuit was filed five minutes after the market closed. On November 8 the board announced an independent review of reports that Huang had been targeted by foreign intelligence, saying it saw no signs of compromised secrets. China's Foreign Ministry called the espionage allegations slander. The company did not respond to requests for comment on the lawsuit.

Even for the biggest enthusiast of SI infrastructure, prepayments for data-center power have turned into non-refundable losses: the bet on compute is getting more expensive and demands ever costlier debt.

SoftBank posts $7 billion loss on write-down of prepayments for US data-center power equipment

Even for the biggest enthusiast of SI infrastructure, prepayments for data-center power have turned into non-refundable losses: the bet on compute is getting more expensive and demands ever costlier debt.

SoftBank Group recorded a net loss of 1.12 trillion yen (about $7 billion) for the July–September quarter, against a profit of 2.5 trillion yen a year earlier, when paper gains came from the rising value of its OpenAI stake. The main cause is a write-down of 597.2 billion yen ($3.7 billion) on advance payments for power-generation equipment for US data-center projects, including the 8 GW PORTS-Pike campus in Pike County, Ohio: procurement plans were revised, $2.85 billion of the prepayment turned out to be non-refundable, and another $880 million went on contract changes and settlements. The company expects rebates of up to $2.1 billion over ten years. Finance costs more than doubled to 461.9 billion yen. LSEG analysts on average had expected a profit of 129 billion yen (per Bloomberg), with forecasts ranging from +233 billion to −113 billion. In total SoftBank has invested $64.6 billion in OpenAI for a stake of about 13%. Despite the loss, its Tokyo shares rose 13.9% in a broad rebound in tech stocks, but they have nearly halved since early June.

Money in SI is shifting from training models to vetting them on real work: human expertise becomes the raw material from which companies build their own vetted agents.

Mercor raises $750 million at a $20 billion valuation to train and vet SI agents for companies

Money in SI is shifting from training models to vetting them on real work: human expertise becomes the raw material from which companies build their own vetted agents.

Mercor, a startup that began as a staffing platform and became a supplier of expert data for training models, has announced a $750 million Series D at a $20 billion valuation — double the $10 billion of October 2025. Benchmark, Felicis, Fundomo, Apollo entities, G Squared, General Catalyst, Glade Brook, Nvidia, Salesforce, Sapphire Ventures, StepStone, XTX Ventures and Zoom participated; part of the sum went to buying shares from employees. The money will go toward evaluating and fine-tuning agents for specific companies: Mercor builds custom benchmarks for each organisation based on its real processes and exceptions. In 2026 the company acquired Deeptune, whose team recreated hundreds of enterprise applications as environments for training agents. Co-CEO Brendan Foody writes that Mercor already spends three times more on inference than on employee salaries. According to Crunchbase, SI startups received $102 billion in the third quarter of 2026 — 64% of global venture capital.

According to AI100, the next leap in SI will depend not simply on even larger clusters but on more efficient algorithms and ways of reasoning. At the same time, society will have to negotiate who controls a technology that has already entered everyday life.

Stanford AI100: the data and compute scaling era runs into limits

According to AI100, the next leap in SI will depend not simply on even larger clusters but on more efficient algorithms and ways of reasoning. At the same time, society will have to negotiate who controls a technology that has already entered everyday life.

The international panel behind Stanford AI100 has released its third report, "AI Goes Mainstream: The Scaling Era and the Rise of Generative AI", its first since 2021. The experts set out 12 findings on the past five years. Growth in data and compute drove the main advances in generative SI, but is now constrained by the shortage of high-quality data and the cost of compute; frontier development is shifting toward reasoning and additional inference-time compute. The report documents research timelines shrinking from years to months or weeks, unprecedented infrastructure investment and pressure on energy and freshwater supplies, and the concentration of frontier development among a small number of US and Chinese players. Other findings address cyber risks, military use, the loss of automatic trust in photographic, audio and video evidence, the public's limited influence over SI deployment, and fragmented international governance. The authors stress that few predicted the technology's current state even five years ago and do not present the report as a precise forecast for the next five.

The state is gaining both mandatory lab reporting and its own technical team to check it: SI safety oversight will be able to rely on independent testing rather than developers' statements alone.

New York opens frontier-lab registration under the RAISE Act and launches independent model-testing network

The state is gaining both mandatory lab reporting and its own technical team to check it: SI safety oversight will be able to rely on independent testing rather than developers' statements alone.

Governor Kathy Hochul has launched the New York Center for AI Responsibility and Research at Binghamton University. More than 80 researchers from institutions across the state will independently evaluate frontier models, develop safety-measurement methods and train specialists using computing resources from the $500 million public-private Empire AI project. The center will work with regulator DIGIT and review developers' published safety frameworks. At the same time, the state opened the registration portal under the RAISE Act. Large frontier-model developers with annual revenue above $500 million must register by January 1, 2027, publish safety frameworks and model-transparency reports, and report critical incidents within 72 hours. Violations carry fines of up to $1 million, rising to $3 million for repeat offenses. As Engadget notes, independent testing is limited to materials companies disclose; the RAISE Act will expand mandatory publications but will not provide access to unreleased models.

The loss of some cloud data is now final: data-center geography alone is no substitute for backups, so Yandex is making backup creation the default setting for every customer.

Yandex acknowledges the permanent loss of 2.8% of Yandex Cloud customer data

The loss of some cloud data is now final: data-center geography alone is no substitute for backups, so Yandex is making backup creation the default setting for every customer.

Yandex said it had restored access to 97.16% of Yandex Cloud customer data following the October attacks on data centers. The remaining 2.8% cannot be recovered because equipment was critically damaged and the affected customers had no backups on the platform. According to the company, customers who used Yandex Cloud's backup tools retained 100% of their data. Yandex has enabled backups by default on the existing terms, though customers can opt out. Engineers continue to distribute compute resources across new locations in different regions of Russia. The strikes on data centers in Sasovo, Kaluga and Vladimir were recorded on October 8, 9 and 11, and restoring key services took most of the month.

Claude is no longer just a conversational assistant and a programmer's helper: it can now manage work during a meeting and immediately turn the discussion into documents, spreadsheets and tasks.

Anthropic turns Claude into an office suite with meetings, spreadsheets, whiteboards and in-chat apps

Claude is no longer just a conversational assistant and a programmer's helper: it can now manage work during a meeting and immediately turn the discussion into documents, spreadsheets and tasks.

Anthropic has introduced a set of office tools. Claude Meetings, in beta for Team and Enterprise plans, joins Google Meet as a visible participant: it follows the conversation, answers questions in the chat, uses connected work apps and drafts plans, tasks and a whiteboard while the meeting is still underway. It identifies itself on the call and does not remain with external guests unless an administrator authorizes it. Claude Whiteboards and Claude Sheets are available in beta on all paid plans: the former turns a group discussion into a structured board, document or presentation, while the latter creates and edits spreadsheets with tabs, formulas and drop-down menus, showing every change. Claude Data, formerly Dashboards, is out of beta and builds dashboards, written narratives and videos from corporate data. Interactive connectors open full Adobe, Atlassian and HubSpot apps alongside the Claude conversation. This is Anthropic's first coordinated push into mainstream non-technical office work, where Microsoft, Google and OpenAI already compete.

Demand for rented SI computing continues to grow faster than supply, but every new data center requires more debt: the industry's central question is shifting from finding customers to the cost of financing.

CoreWeave raises its annual forecast after quarterly revenue grows 168% to $3.66 billion

Demand for rented SI computing continues to grow faster than supply, but every new data center requires more debt: the industry's central question is shifting from finding customers to the cost of financing.

CoreWeave's third-quarter revenue reached $3.66 billion, compared with the LSEG consensus of $3.55 billion and $1.36 billion a year earlier. Adjusted loss per share was $0.95 versus the expected $1.16, but net loss grew from $110 million to $709 million. The company raised its 2026 revenue forecast to $13–13.3 billion and adjusted operating income forecast to $1.275–1.375 billion. Its revenue backlog increased from $104 billion in June to $132.5 billion. At the same time, debt reached almost $47 billion and planned annual capital expenditures are $37–39 billion. CFO Nitin Agrawal said the parent company would not issue new unsecured debt through the end of 2027 at current yields and would rely primarily on asset-backed financing and customer prepayments. Management retained its target of 1.85 GW of active power by the end of 2026. Shares rose in after-hours trading, despite losing more than 40% of their value over the previous month.

Sources 1

Checking a model's behavior is not enough: the weights themselves will need to be protected and compared with the original, just like signed software code, especially when an agent is allowed to execute commands.

Microsoft shows how altering fewer than ten neurons inserts a hidden command into an SI agent

Checking a model's behavior is not enough: the weights themselves will need to be protected and compared with the original, just like signed software code, especially when an agent is allowed to execute commands.

Microsoft researchers have described Lazarus, a method that reprograms fewer than ten rarely active neurons in a language model's weights without retraining. The model retains normal behavior and passes standard evaluations but emits a malicious token sequence when a hidden trigger appears. A neuron is trained to recognize the trigger on roughly 50 examples; a chain of such modifications can form a command to download and run code, which the agent executes with its own terminal and tool permissions. The trigger can be an ordinary event such as the arrival of 2027. The attack requires write access to the weights file, so the risk affects not only open models but also internal systems if storage or a pipeline is compromised. Microsoft Foundry already ties open weights to trusted publishers and scans their environment; the company is developing weight scanning with frontier agents, red-team checks and cryptographic signing of model provenance. The researchers emphasize that scanning cannot prove safety: reliable comparison requires the original unmodified weights.

As models become more capable, their internal explanations become a less reliable window into their intentions, so agent safety will require checking actions and developing new monitoring methods.

OpenAI finds that base-model improvements and reinforcement learning weaken oversight of chain of thought

As models become more capable, their internal explanations become a less reliable window into their intentions, so agent safety will require checking actions and developing new monitoring methods.

OpenAI has published research into factors affecting the ability to monitor a model's intentions through its chain of thought (CoT). Across a series of internal and released models, the ability to reason without CoT grew smoothly as base-model perplexity fell; architecture choices strengthened this effect only on some planning and sequential-computation tasks. Reinforcement learning increased CoT controllability — the model's ability to add, remove or alter reasoning tokens on request — even without evaluating the reasoning process itself; tool access and supervision added a small effect. This controllability was also passed to the next generation through fine-tuning on a teacher model's responses. OpenAI treats both metrics as approximate indicators of future monitor evasion but explicitly notes their limitations: in a controlled pair of models, greater CoT controllability did not produce greater evasion success. The company plans to develop direct behavioral evaluations and other oversight methods rather than relying solely on reading reasoning.

An SI app builder's share button can publish not only the app interface but also embedded working data; secure links and privacy by default are becoming part of protecting corporate secrets.

Researcher finds more than 35,000 publicly accessible links to Grok Build projects containing confidential data

An SI app builder's share button can publish not only the app interface but also embedded working data; secure links and privacy by default are becoming part of protecting corporate secrets.

Researcher Aprozper of Claroc found that addresses of published Grok Build applications appear in passive DNS records, even though the random portion of the link was supposed to restrict access to people who know it. In a sample of 43,601 links, more than 35,000 opened the application and its associated data to any visitor. The researcher reported finding internal business plans, advertising proposals, medical examination results, and employee names and ID numbers; 24 applications belonged to SpaceXSI itself. In his tests, Grok Build defaulted to "anyone with the link," and users had to select "invited users" or "only you" to restrict access. A representative of Cursor, which belongs to SpaceXSI, told Ars Technica that apps are private by default and become public only at the creator's choice; Aprozper's test results contradict this. Other app builders pose similar risks, whereas Claude, ChatGPT and Gemini place the random part of a public link in the URL path, which DNS does not see.

Demand for OpenAI's products continues to turn rapidly into revenue: quarterly receipts have reached a scale that gives the company more resources for costly model and infrastructure development.

OpenAI's quarterly revenue reaches $10 billion for the first time after almost tripling in a year

Demand for OpenAI's products continues to turn rapidly into revenue: quarterly receipts have reached a scale that gives the company more resources for costly model and infrastructure development.

According to Bloomberg, OpenAI's revenue for the three months through the end of September reached $10 billion for the first time: almost 50% more than in the previous quarter and nearly three times the year-earlier figure. The agency cites a person familiar with the matter who spoke on condition of anonymity because the financial information has not been made public. According to that person, growth was driven by demand for OpenAI's latest software, and the figures were presented to investors during a call on Tuesday.

Medical SI is moving beyond answering questions to coordinated work across the patient's record, insurer and clinical guidelines: physicians will be able to spend less time finding and transferring data and more time with patients.

Google launches Gemini for Healthcare for medical records, insurance rules and authorizations

Medical SI is moving beyond answering questions to coordinated work across the patient's record, insurer and clinical guidelines: physicians will be able to spend less time finding and transferring data and more time with patients.

Google Cloud has released Gemini for Healthcare in preview with partners including HCA Healthcare, Highmark Health, Northwell Health, Optum and Stanford School of Medicine. The system combines data from medical records, billing, insurance coverage, Medicare and clinical guidelines from ASCO and the BMJ. Five ready-made skills are available at launch: synthesizing a patient's history, preparing for appointments, reviewing coverage and treatment authorizations, estimating costs and other multistep processes. Epic and Oracle Health connectors read records without permission to modify them, and conclusions include source citations so the physician can verify every fact and make the decision. The agent acts within defined guardrails and the employee's role-based permissions; data remains within the organization's cloud environment and Google does not use it to train foundation models. According to Google Cloud and Harris Poll, documentation and administrative work consumes nearly 28 hours of clinicians' time each week; the product is intended to return that time to patient care.

Detailed high-resolution image generation and editing can now run on local hardware without dependence on a closed API.

MiniMax opens H3-Image weights for image generation and editing at up to 16 megapixels

Detailed high-resolution image generation and editing can now run on local hardware without dependence on a closed API.

H3-Image combines text-to-image generation and editing from one or more references in a single checkpoint. The model directly generates images of about 16 MP without an upscaler, including 4096×4096 and aspect ratios up to 21:9, and accepts up to 16 references. The distilled version uses six sampling steps; its diffusion transformer has 33.1 billion parameters. Weights and configurations for SGLang, vLLM-Omni, diffusers and ComfyUI are published on Hugging Face, with support for Nvidia and AMD accelerators. MiniMax claims 4K generation in under 10 seconds; speed depends on the hardware configuration. Local use also requires the MiniMax-H3 text encoder, and the company offers a separate model for expanding short prompts.

The market is weighing not only the CEO scandal but also governance, export-control and key-person risks at Nvidia: two days of news about its chief executive have turned into hundreds of billions of dollars in market-value losses.

Nvidia heads for more than $470 billion in market-value losses over two days of falling shares

The market is weighing not only the CEO scandal but also governance, export-control and key-person risks at Nvidia: two days of news about its chief executive have turned into hundreds of billions of dollars in market-value losses.

According to The Korea Times, Nvidia shares fell more than 6% on Tuesday after a 4.3% decline on Monday, putting the company on track to lose more than $470 billion in market value over two sessions. Microsoft overtook Nvidia in market value for the first time since June 2025, becoming the world's second-most valuable company. The sell-off continues amid the board's investigation and the FBI's inquiry into Jensen Huang's connection to a woman whom US intelligence agencies suspect of working for Chinese intelligence; Bank of America analysts caution against equating the allegations with proven damage to the business.

SI accelerators are beginning to specialise: instead of one universal chip, variants are appearing for protecting data while in use and for huge contexts, which makes the alternative to GPUs more practical.

Google launches two new TPUs: TPU 8tx, designed with AMD for confidential computing, and TPU 8i Max with double the memory

SI accelerators are beginning to specialise: instead of one universal chip, variants are appearing for protecting data while in use and for huge contexts, which makes the alternative to GPUs more practical.

Google has expanded the eighth generation of its tensor processors with two variants. TPU 8tx is built on the same silicon as TPU 8t but with hosts on fifth-generation AMD EPYC processors: it is designed for training and running models on confidential data and protects data and model weights at the hardware level while they are in use, and also allows fine-tuning small models without moving data off the device. TPU 8i Max combines two TPU 8i inference chips in a single package, doubling compute, SRAM and memory bandwidth and fitting twice as much silicon into the same rack. This is already the fourth TPU system this year, Google says, and the number of TPU customers is approaching one hundred, including frontier labs and trading firms. The specifications come from Google's claims.

Security monitoring is going mobile: the camera no longer stares at a single point but takes to the air or follows a route on wheels — autonomous robots are moving from demos to products with announced prices and shipping dates.

Amazon unveils a $250 Ring drone camera and a $1,000 wheeled patrol robot

Security monitoring is going mobile: the camera no longer stares at a single point but takes to the air or follows a route on wheels — autonomous robots are moving from demos to products with announced prices and shipping dates.

Ring (an Amazon unit) has launched Ring Robotics with two devices. Ring Drone Cam is a $250 flying security camera that takes off on command and circles the house, including the second floor, then returns to its base to recharge. Ring Rover Cam is a $1,000 wheeled patrol robot for businesses that moves along a set route and films what is happening. Both devices use SI processing of footage to distinguish urgent events from ordinary activity. Rover Cam shipments are promised for December 16, 2026, and Drone Cam shipments for January 27, 2027. At the same time Ring unveiled a smart lock and an expanded line of cameras.

Creative apps are becoming agent-controllable: bots take over routine work like "make 50 versions of a document", leaving the human to review and choose — the profession is gradually shifting from execution to setting the task and curating.

Adobe will open Photoshop, Illustrator and Premiere to third-party SI agents via local MCP and add "always-on" agents

Creative apps are becoming agent-controllable: bots take over routine work like "make 50 versions of a document", leaving the human to review and choose — the profession is gradually shifting from execution to setting the task and curating.

At the Adobe MAX conference in Miami, Adobe announced that third-party developer agents (OpenAI, Google, Anthropic) will be able to work with files directly on the user's computer through local integrations over the open MCP protocol and modify layers, vectors, sequences and bins while keeping everything hand-editable. The feature will enter closed beta in early 2027; in a demo, ChatGPT produced nine editable poster layouts from a single file in 9 minutes. Adobe also showed "always-on" agents for recurring tasks (sorting assets, quality checks) and updates to Firefly AI Assistant: a Plan Mode with plan approval, model selection and binding to brand guidelines. Adobe's creative agent is already built into Premiere, Photoshop, Illustrator, InDesign, After Effects, Frame.io and the mobile version of Premiere; in Premiere it can apply effects to all clips in a sequence and track a masked object across frames. At the same time Google and Adobe announced the integration of Adobe tools into Google apps.

Language-model-driven robots completed real factory trials with 96% success, and alongside growing capabilities comes a common way to measure how hard they are to steer toward dangerous actions.

Google DeepMind releases Gemini Robotics 2.5 and open ASIMOV-Jailbreak safety benchmark

Language-model-driven robots completed real factory trials with 96% success, and alongside growing capabilities comes a common way to measure how hard they are to steer toward dangerous actions.

DeepMind has released a new version of its robot control system, Gemini Robotics 2.5. According to the company, task success rates rose 30–100% versus the previous version and speed by 20–30%, while the model is noticeably better at small, precise movements: the Apptronik Apollo 2 humanoid ties a bow, and the model handles serving food with tongs and working with a pipette. In trials outside the lab, Apollo 2 robots at Mercedes-Benz plants in Europe assembled auto parts with 96% success — which is still not enough for industrial use, DeepMind notes. The simulation environment for training was generated with Gemini 4 Argon, but the model itself runs on cloud servers, meaning there is latency. For safety testing, an open set called ASIMOV-Jailbreak was released: 1,500 prompts that provoke a robot into unsafe actions; with Argon in the loop, the system blocks 98.8% of attacks. The model is available to partners — Agile Robots, Apptronik and Boston Dynamics.

Whether you can hand a purchase to an agent hinges not on how capable the models are but on shop doors being closed: while marketplaces block bots, "agentic commerce" remains a promise, not a habit.

eBay starts blocking third-party SI agents, including Meta's Muse

Whether you can hand a purchase to an agent hinges not on how capable the models are but on shop doors being closed: while marketplaces block bots, "agentic commerce" remains a promise, not a habit.

A Verge reader used Meta's Muse agent for weeks to compile daily watch roundups on eBay, and over the weekend Muse stopped pulling links and listing details, reporting a block; The Verge reproduced this with other queries. eBay had earlier changed its user agreement to ban "agents that buy on a person's behalf" and LLM bots that place orders without human review, but until recently Muse was still able to browse listings. An eBay spokesperson said the platform takes a selective approach to third-party access and restricts agents partly because of unauthorised scraping and training on its data. Retailers' policies diverge: Amazon blocks external agents entirely, Target allows them in some cases and blocks them in others, and Shopify has opened a catalogue of a billion products for agentic shopping. At the same time, agent usage is still tiny: Uber completed four rides through third-party agents in the third quarter, and DoorDash "several hundred" orders against 993 million orders in the quarter.

Sources 1

Investors are backing the transition of industrial robots from single operations to general-purpose workers: a model already trained for difficult welding will be transferred to new tasks and robotic platforms.

Path Robotics raises $300 million at a $2.45 billion valuation for manufacturing robots

Investors are backing the transition of industrial robots from single operations to general-purpose workers: a model already trained for difficult welding will be transferred to new tasks and robotic platforms.

Path Robotics has closed a $300 million Series E round at a post-money valuation of $2.45 billion, led by Georgian, whose Emily Walsh will join the board. Addition, Drive Capital, IronArc Ventures, Tiger Global, Yamaha Motor Ventures, D. E. Shaw Ventures and NAVER Ventures also participated. The company has raised more than $610 million since 2018. According to the company, annual recurring revenue grew 17-fold over 18 months, while headcount increased 97% over the past year. Its Obsidian model locates weld joints, plans welds and adapts to conditions instead of following a rigid program. The funding will support new manufacturing operations and robot form factors, compute and model training. Customers include HII, Lockheed Martin and Saronic; earlier this year, Path signed a $600 million production agreement with HII for US Navy shipbuilding programs. The mobile Rove system brings Obsidian to a quadruped platform so large parts can be welded where they stand.

SI modernization at a major company brings not only new tools but also the transfer of hundreds of familiar office roles to a contractor: genuine retraining, internal transfers and severance terms now matter directly to workers.

Qantas to cut 850 corporate jobs in Australia as some work moves to Accenture

SI modernization at a major company brings not only new tools but also the transfer of hundreds of familiar office roles to a contractor: genuine retraining, internal transfers and severance terms now matter directly to workers.

Qantas has signed a multi-year agreement with Accenture to modernize internal systems, processes and data and accelerate its use of SI. About 850 corporate positions in Australia are set to be cut or transferred to the external provider, with the first changes due by mid-2027. Finance, procurement, people and marketing will be affected first, with some services moving primarily to Accenture teams in India and partly to teams in Australia. Qantas CEO Vanessa Hudson linked the decision to more than 1,000 legacy applications that were never designed to work together and promised consultations with employees and unions, retraining and transfers to other roles. The Australian Services Union demanded an explanation for the contradiction with the company's August statement that it had no plans to move jobs offshore, and called for a concrete retraining and compensation plan. According to The Guardian, announcements by NAB, Canva and Qantas bring the total number of corporate positions targeted for cuts, relocation or restructuring in Australia over the past month to about 4,550.

ChatGPT is becoming not only a place to search for products but also a storefront where a company can receive a customer through its own verified agent, reducing the path from question to purchase to a single conversation.

OpenAI launches verified business profiles and companies' own Business Agents in ChatGPT

ChatGPT is becoming not only a place to search for products but also a storefront where a company can receive a customer through its own verified agent, reducing the path from question to purchase to a single conversation.

According to OpenAI's description, the company has begun beta-testing Businesses in ChatGPT. Through Business Hub, a company can verify its organization, update its details and connect a product catalog with prices and availability; Shopify merchants can connect directly from their admin panel. The company also sees aggregated metrics for mentions, website clicks and conversation topics in which ChatGPT cites it, but does not receive user conversations. Eligible US online retailers and service providers can create a Business Agent: when such a company appears in a ChatGPT response, the user can open its profile and enter a separate, clearly labeled conversation with the brand's agent. The agent uses the company's instructions and materials to answer questions and help compare offerings. OpenAI links the new layer to ChatGPT's 1.2 billion weekly users.

The smart ring is evolving from a way to monitor the body into an entry point for medical care: a connected system of SI and physicians will be able to take the wearable's measurements into account.

Oura agrees to acquire healthcare platform Counsel Health for $625 million

The smart ring is evolving from a way to monitor the body into an entry point for medical care: a connected system of SI and physicians will be able to take the wearable's measurements into account.

Smart-ring maker Oura has agreed to acquire Counsel Health, a primary-care service in which medical SI communicates with the user and a licensed physician joins when needed. The service has been integrated into the Oura app since June. Following the deal, the companies plan to combine physiological data from the ring with consultations, prescriptions, lab results, chronic-disease monitoring and specialist referrals. Counsel CEO Muthu Alagappan will become Oura's chief clinical SI officer, and his 70-person team will join the company. Counsel will continue serving other enterprise partners. The startup previously raised a $25 million Series A led by Andreessen Horowitz and GV. The deal is expected to close before the end of 2027. Oura also reported quarterly revenue growth of 153% to $533 million.

The review has moved from a general board statement to an investigation involving outside lawyers and a former federal prosecutor: its findings will affect Nvidia's governance and the assessment of risks surrounding access to its technology.

Nvidia board assigns former federal prosecutor to investigate circumstances of Jensen Huang's relationship

The review has moved from a general board statement to an investigation involving outside lawyers and a former federal prosecutor: its findings will affect Nvidia's governance and the assessment of risks surrounding access to its technology.

Nvidia's independent directors have hired law firm Wachtell, Lipton, Rosen & Katz and Joon Kim, former acting US Attorney for the Southern District of New York, to conduct an independent investigation into CEO Jensen Huang's relationship with Alice Huang, whom US intelligence agencies suspect of working for Chinese intelligence. The review continues the investigation announced on November 8 into a possible foreign-influence effort and risks to company secrets and national security. The board previously said it had no reason to believe such secrets had been compromised. Jensen Huang acknowledged the extramarital relationship but, according to US President Donald Trump, said he had not known about the suspicions. China's Foreign Ministry called the allegations an attempt to smear China. According to the BBC, leaders of the House Foreign Affairs Committee have also asked Huang to attend a closed-door briefing next week.

Servers for SI are gaining not only faster processors but also a way to work around the DRAM shortage: access prediction allows inexpensive flash memory to serve as part of RAM without a proportional loss of speed.

First independent EPYC Venice tests show gains of up to 93% over Zen 5, while MEXT replaces some DRAM with flash

Servers for SI are gaining not only faster processors but also a way to work around the DRAM shortage: access prediction allows inexpensive flash memory to serve as part of RAM without a proportional loss of speed.

Tom's Hardware tested dual-socket servers with AMD EPYC 9996 and EPYC 9686F processors based on Zen 6. In the multithreaded integer SPEC CPU 2026 test, a system with two 256-core EPYC 9996 chips was 93% faster than the previous flagship EPYC 9965, while floating-point tests showed gains of roughly twofold over Zen 5; these are results for specific workloads, not a general improvement across all tasks. The compute chiplets are manufactured on TSMC's 2nm N2P process, and the platform supports up to 256 cores, 16 memory channels and PCIe 6.0. AMD also introduced MEXT: a Linux driver and an SI engine running on one EPYC core predict data accesses, keep active pages in DRAM and move the rest to an NVMe drive. According to AMD, cutting DRAM by 50% in electronic-design workloads preserved 95% of performance, while having 25% of the memory available preserved 80%; in some scenarios, the company promises up to 3.5 times more performance per dollar of memory. AMD is already deploying MEXT in its own server infrastructure.

Naval surveillance is shifting toward a persistent network of robots and sensors working alongside crewed ships: autonomous systems will be able to cover areas for longer where continuous human patrols are too expensive.

UK allocates £90 million to Atlantic Net autonomous-system network for underwater surveillance

Naval surveillance is shifting toward a persistent network of robots and sensors working alongside crewed ships: autonomous systems will be able to cover areas for longer where continuous human patrols are too expensive.

The UK Ministry of Defence has signed a framework agreement with Ocean Infinity Solutions for Atlantic Net, the first phase of the Atlantic Bastion program. Under the ministry's plan, a network of autonomous and lightly crewed platforms, underwater sensors, SI and communications systems will conduct persistent surveillance in the North Atlantic and provide ships and Astute-class submarines with earlier warnings of submarines and other activity. The program's objectives include detecting Russian submarines and protecting communications cables and energy infrastructure. The £90 million investment will support almost 200 skilled jobs at Ocean Infinity and its suppliers, according to the ministry. The system will be developed with NATO and the Joint Expeditionary Force; the ministry did not give an expected date for entry into service.

Before granting an agent access, its organization can be verified through an open standard rather than incompatible closed directories.

Linux Foundation launches ANS, an open standard for verifiable SI-agent identities

Before granting an agent access, its organization can be verified through an open standard rather than incompatible closed directories.

Agent Name Service uses existing DNS infrastructure and certificates to register agents, publish their organizational ownership, discover them and verify identity before granting access to systems. The seven-document specification covers domain-based naming, registration and versioning, DNS publication, public-log auditing, change monitoring and mutual agent authentication. The project has moved under the neutral governance of a Linux Foundation technical committee. It is supported by Cisco, Cloudflare, DigiCert, GoDaddy, Infoblox, Salesforce, OWASP and four other organizations. AGNTCY Agent Directory Service already verifies ANS identifiers, and ANS Finder is compatible with the Agentic Resource Discovery specification.

Voice agents and meeting transcription will become more accurate in multilingual conversations, while pauses and speaker changes will be recognized more naturally.

ElevenLabs releases Scribe v3 and Scribe v3 Live for recorded and streaming transcription

Voice agents and meeting transcription will become more accurate in multilingual conversations, while pauses and speaker changes will be recognized more naturally.

In ElevenLabs' internal tests across 102 languages, Scribe v3 reduces word errors by 27% compared with v2, while Scribe v3 Live reduces errors by 43% relative to v2 Realtime. The streaming version gains speaker labels for the first time and uses conversational context to detect the end of a turn, avoiding interruptions during long pauses. For recorded audio, diarization errors fell by 12% overall and by 17% with three or more speakers. The models support key-term lists, formatting rules written in natural language and language switching within a sentence. Scribe v3 Live is already available in ElevenAgents, and both models are available through the API; some results were measured on an internal ElevenLabs dataset that the company promises to open-source.

Instead of manually reading dozens of threads, shoppers will get a summary of real experiences with links to the originals: product discovery is beginning to draw on live discussions rather than only store descriptions and review sites.

Reddit turns user discussions into SI reviews of 10,000 products

Instead of manually reading dozens of threads, shoppers will get a summary of real experiences with links to the originals: product discovery is beginning to draw on live discussions rather than only store descriptions and review sites.

Reddit has begun testing Reddit Reviews: searching for a product or category opens a page where SI gathers relevant discussions, summarizes opinions, and shows the share of positive, mixed and negative conversations alongside recurring pros and cons. Users can filter comments by topic, open the original discussion, ask a follow-up question, and see products that users frequently compare. If no review is available yet, Reddit AI Search answers the question using material from the communities. The test will roll out to US users on iOS, Android and the web over the next few weeks; the initial selection covers about 10,000 products in electronics, health and beauty, home and garden, and fashion and apparel. Companies cannot pay to have a product included; some images, prices and purchase links come from brand and retailer catalogs. In 2027, Reddit plans to add more products and expand beyond the US.

Tax SI is moving from answering questions to preparing finished work with verifiable references and steps: professionals can assign an entire document to agents while retaining judgment and decision-making.

Blue J raises $176 million for SI agents that complete tax work end to end

Tax SI is moving from answering questions to preparing finished work with verifiable references and steps: professionals can assign an entire document to agents while retaining judgment and decision-making.

Toronto-based Blue J has raised $176 million in a Series E round led by JMI Equity at a valuation of more than $700 million. Oak HC/FT, Sapphire Ventures and Intrepid Growth Partners also participated, and JMI's Jack Duane will join the board. In the 15 months since Series D, annual recurring revenue grew 165%, and the user base reached 6,500 organizations and 45,000 tax professionals. The company's research system has processed more than 14 million tax questions. Blue J is now developing Workflows: agents review tax returns, prepare responses to IRS notices, analyze partnership, LLC and trust agreements, assess tax obligations across different states, and draft client emails. The result includes sources and a record of the steps taken so a professional can review the work and make the decision. Part of the capital will fund expansion into the UK and other countries.

An autonomous agent can be allowed to do more when dangerous steps are intercepted before execution and assessed against the task: control shifts from reviewing logs after a failure to preventing the failure itself.

Bay emerges from stealth with $50 million to vet SI agent actions before execution

An autonomous agent can be allowed to do more when dangerous steps are intercepted before execution and assessed against the task: control shifts from reviewing logs after a failure to preventing the failure itself.

Israeli company Bay has raised $50 million in the nine months since its founding: $37 million in a Series A led by Battery Ventures and a previously undisclosed $13 million seed round led by YL Ventures and Striker. Micron Ventures, M12 and Wiz CEO Assaf Rappaport also participated in the Series A. Before an action is executed, the platform compares it with the agent's assigned task and prior session context, then permits it, blocks it or refers it to a person. The analysis covers prompts, responses, tool calls and add-ons; rules run on the device, in the browser, in a cloud workspace or in SaaS without installing another full endpoint agent. According to co-founder Yuval Zacharia, deployment across 150,000 Micron endpoints took less than five minutes. The company reports dozens of customers, including Wix, ClickUp and Micron, and will use the funds for product development and sales in the US and Europe.

Generative SI will be applied not only to medicines but also to crop protection: computers will screen billions of possible molecules, and experimenters will test which actually work and are safer than existing herbicides.

MIT to receive up to $53 million for generative SI to develop new herbicides

Generative SI will be applied not only to medicines but also to crop protection: computers will screen billions of possible molecules, and experimenters will test which actually work and are safer than existing herbicides.

An MIT-led team will receive up to $53 million from ARPA-H for the PLOW project, which seeks alternatives to modern herbicides. Generative models will propose chemical molecules and small interfering RNAs that act on the same plant enzymes as glyphosate while accounting for synthesis cost, the likelihood of weed resistance, and risks to people and the environment. The models are expected to generate billions of candidates; a second group will conduct high-throughput screening and feed the results back to improve the models, while a third will test selected substances in greenhouses. The first greenhouse tests of initial RNAi candidates are expected by the end of the program's first year. A separate track will develop microbial systems that break down agrochemicals before they reach food, soil and water. The project is at the research stage, not a finished product.

Sources 1

Restrictions on access to advanced computing could be enshrined not only in administration decisions but also in defense legislation: that would make them harder to change and build location tracking into the technical infrastructure of chip exports.

Jeffries backs including SI chip export restrictions in the US defense budget

Restrictions on access to advanced computing could be enshrined not only in administration decisions but also in defense legislation: that would make them harder to change and build location tracking into the technical infrastructure of chip exports.

House Democratic Minority Leader Hakeem Jeffries said he would work to preserve three provisions from the Senate version of the annual National Defense Authorization Act in the final bill: the AI OVERWATCH Act, the MATCH Act and the Chip Security Act. They would allow Congress to block exports of advanced SI chips to countries the US considers foreign adversaries, restrict sales of essential chipmaking equipment to listed countries of concern, and require a location-verification mechanism for restricted chips. Jeffries spoke after a classified briefing with administration officials and called for a separate detailed briefing on the situation involving Nvidia and its implications for national security. No position from the Chinese authorities specifically on the new discussion of these provisions was found in the verified sources.

Sources 1

OpenAI has acquired not a ready-made service but a team experienced in building a travel agent: trip planning could become one of the specialized capabilities of personal SI.

OpenAI acquires travel agent Odessia and brings in Sonder founder Francis Davidson

OpenAI has acquired not a ready-made service but a team experienced in building a travel agent: trip planning could become one of the specialized capabilities of personal SI.

OpenAI has confirmed its acquisition of Odessia and the move of founder Francis Davidson to a team that will work on the future of travel in the company's products. Davidson previously founded accommodation service Sonder and created Odessia less than a year ago. The startup raised $6 million in pre-seed funding led by Sequoia and launched its agent about five months ago; it built a bookable itinerary around a user's preferences in minutes. According to Davidson, tens of thousands of people used the service over the summer to plan travel worth $60 million on an annualized basis, while he built the company by directing teams of SI agents. Odessia shut down its service after the sale. The parties did not disclose the deal terms or OpenAI's new travel products.

The proposal is to build SI into the operating system itself rather than another application: different models could share local memory and computing resources while the user manages their access from one place.

Cyanogen creators unveil Kora, a Linux OS with shared memory for different SI models and agents

The proposal is to build SI into the operating system itself rather than another application: different models could share local memory and computing resources while the user manages their access from one place.

Kora Systems, a team with backgrounds at Cyanogen, Android and Playtron, has shown its Linux-based Kora desktop operating system. Instead of isolated chats and tabs, it connects models and agents at the system level: the local Cortex engine distributes work across the CPU, GPU, NPU and other machines on the network, while cloud frontier models handle heavier tasks. A neural file system searches documents by meaning, and workspaces restrict each agent's access to files, devices and services; context and memory remain on the computer by default. Kora supports x86-64 and Snapdragon X2 Elite, a compatibility layer runs some Windows applications, and GameFrame retains GameOS's gaming mode. Qualcomm and Kora demonstrated a local 26-billion-parameter model on Snapdragon X2 Elite. The first computers with the enterprise HUMAIN OS built on this foundation are expected in 2027, with standalone Kora following later that year.

Agents no longer need to compose text for classification and routing: an application can receive a typed decision and its probability directly, faster and in a consistent format across different models.

Hugging Face introduces shared decision-model interface with probabilities instead of generated text

Agents no longer need to compose text for classification and routing: an application can receive a typed decision and its probability directly, faster and in a consistent format across different models.

Hugging Face has standardized support for decision models, which evaluate permitted decisions directly rather than writing a response token by token. A request consists of context and typed questions: choice selects an option and returns a probability distribution, score calculates an expected value on an ordered scale, and noul gives the probability of a yes-or-no answer. This allows one call to route a request, assess urgency and check a condition without parsing potentially malformed JSON or paying to generate an explanation; the API example returns zero output tokens. Transformers has gained DecisionPipeline, AutoModelForDecision and DecisionConfig, while Transformers and llama.cpp support the shared /v1/systemone endpoint. The compatible interface works with purpose-trained models and with ordinary autoregressive LLMs, whose option logits are normalized without text generation. The Hub has added a dedicated tag for finding such models, and the Trainer API already supports fine-tuning them.

SI is changing not only which tasks clients give freelancers but also the structure of projects: instead of one narrow service, clients increasingly ask for several skills to be combined into a larger job.

OpenAI and Upwork find SI freelance projects are 45% more diverse and combine larger tasks

SI is changing not only which tasks clients give freelancers but also the structure of projects: instead of one narrow service, clients increasingly ask for several skills to be combined into a larger job.

According to OpenAI's description, its joint research with Upwork compared categories of ChatGPT use with freelance job postings. Across 63 matched categories, 43 areas that were relatively common in ChatGPT lost share on Upwork while less common areas gained share; the remaining 20 showed the reverse pattern, so the authors do not reduce changing demand to simple substitution of humans by models. From early 2023 to late 2025, the diversity of tasks within SI-related postings grew by 45%, from the equivalent of 14.6 roughly equally common tasks to 21.2. Clients more often combined types of work that had rarely appeared together, such as video generation, editing and animation in one project. Hourly rates were similar for SI-related and other projects, but fixed-price budgets and hours were larger for SI-related work, consistent with a broader scope. The research concerns the freelance market and does not establish the same relationship for employment overall.

Sources 1

Autonomous delivery is becoming a routine urban service in another city of over a million people: scale now depends on the number of connected warehouses and districts rather than individual robot demonstrations.

Yandex launches robot delivery in Yekaterinburg

Autonomous delivery is becoming a routine urban service in another city of over a million people: scale now depends on the number of connected warehouses and districts rather than individual robot demonstrations.

Yandex autonomous rovers have begun delivering orders in selected locations in Yekaterinburg's Oktyabrsky, Chkalovsky and Leninsky districts. Six Yandex.Lavka dark stores are connected at launch. Robot delivery will later expand to four more districts, and the number of connected dark stores is planned to reach 20. The service already operates in more than ten cities, including Moscow, Saint Petersburg, Kazan, Nizhny Novgorod and Tyumen. In April, the company said it aimed to deploy 5,000 rovers across all cities where it operates by the end of 2026 and 20,000 within two years.

Sources 1

Work conversations become data for a personal agent without having to hand them to an external service: a company can use its own model while preserving each employee's existing access rights.

T-Technologies adds an SI assistant and MCP server to its Time corporate messenger

Work conversations become data for a personal agent without having to hand them to an external service: a company can use its own model while preserving each employee's existing access rights.

The Time corporate messenger has gained a built-in SI assistant, discussion summaries and an MCP server for connecting external agents. The assistant finds messages and missed questions, gathers context, reviews open tasks and writes on the user's behalf; a separate function summarizes a thread or channel with filters by employee and period. Through MCP, a third-party agent can read conversations and act only within the individual user's permissions. The SI functions do not access the internet and work through internal interfaces; a company can connect a model running in its own environment. According to T-Bank, employees sent more than one million unique queries during three months of internal testing.