Show only each day's banner story

OpenAI releases GPT-6 Cyber for vulnerability research and limits access to vetted researchers

Finding and validating vulnerabilities is becoming much faster, but strict access and control over which systems a model may investigate are becoming equally important.

The model is available through the API and Codex to an initial group of approved Daybreak Red customers authorized to develop exploits and conduct authorized testing. OpenAI classifies GPT-6 Cyber as Critical under its Preparedness Framework. It scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber; in the controlled RemoteExploitBench, it completed 95% of tasks involving remote exploitation of known vulnerabilities. According to OpenAI, the system discovered 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity and produced working exploits for 948 privilege-escalation vulnerabilities in the Linux kernel. In a test featuring tempting unrelated services, the model stayed within the authorized scope in 99.93% of cases without a deployment prompt and in every test with the prompt. Access includes monitoring, network restrictions and additional rules for critical infrastructure.

Building compute capacity is becoming not only a race for energy but also a commitment to the people living near data centers.

OpenAI commits to covering its data centers' energy and infrastructure costs and publishing water use

Building compute capacity is becoming not only a race for energy but also a commitment to the people living near data centers.

According to OpenAI and CNBC, the company has published a National Community Compact and Compute Campus Policy Blueprint for future compute campuses it will own and operate. It promises to pay the full cost of its electricity and the necessary grid infrastructure, provide financial assurances before major spending in case a project is scaled back or cancelled, and publish construction plans plus annual figures for water, energy, noise and hiring. Each site will have a water plan prioritising closed-loop or low-water cooling. OpenAI will also stop using non-disclosure agreements with public officials, except for narrow protections covering security and trade secrets. A separate blueprint asks policymakers to apply similar requirements across the industry: special tariffs for large electricity users, public terms for tax incentives and clawbacks when commitments are not met. The announcement followed local US elections in which, according to CNBC's count, voters backed restrictions or rejected pro-development measures in 40 of 41 cases.

Autonomous delivery is moving from isolated pilots to a large network: capital and hundreds of new sites will test its economics at national scale.

Zipline raises $1.5 billion at a $21.5 billion valuation and expands drone delivery to 300 more Walmart sites

Autonomous delivery is moving from isolated pilots to a large network: capital and hundreds of new sites will test its economics at national scale.

Paradigm, WCM Investment Management and Layer Global led Zipline's round, with Uber, Fidelity, Tiger Global, J.P. Morgan, ARK Invest and Salesforce Ventures also participating. The capital will expand US manufacturing, hiring and a nationwide autonomous delivery network. Under a new multi-year agreement, Walmart and Zipline plan to launch another 300 sites, while Uber is targeting one million deliveries per day with Zipline by 2029. According to the company, more than one million deliveries have already been completed in Dallas–Fort Worth in 2026, and the median October order arrived 19 minutes after checkout. Among a group of long-standing restaurant partners, Zipline's share of delivery revenue rose from 10.8% in March to 34.9% in October. Their delivery gross merchandise value increased by 26%, with 80% of that growth coming from drone orders.

A rare internal assessment points to a feedback loop in which SI already helps build subsequent models faster, although the size of the effect remains highly uncertain.

METR estimates Anthropic's model development has roughly doubled in pace with help from SI

A rare internal assessment points to a feedback loop in which SI already helps build subsequent models faster, although the size of the effect remains highly uncertain.

From August through October, METR examined Anthropic's internal metrics, interviews and data on pretraining, reinforcement learning and employees' use of SI. On the AECI capability index, model scores have grown 1.9 times faster since October 2025 than at the previous rate; the estimate's 95% confidence interval ranges from 1.5 to 2.8. METR considers it likely that most of the current acceleration comes from using SI in research and development, but emphasizes substantial uncertainty: a plausible range for overall acceleration extends from roughly 1.3 to 3 times. The researchers found no convincing bottlenecks in data, algorithms or compute that would necessarily halt further acceleration over the next year, but do not consider that conclusion definitive. Anthropic had the right to redact confidential information and removed most of the evidence METR described as most informative. Outside readers must therefore rely more heavily on the authors' judgment rather than independently examining the full body of data.

Sources 1

A major new computing project ties SI expansion to dedicated power generation, lower water use and commitments to the local community.

Google will invest $17 billion in its first Wyoming data center and an energy system with carbon capture

A major new computing project ties SI expansion to dedicated power generation, lower water use and commitments to the local community.

Google is developing the Laramie County project with Tallgrass Energy, Black Hills Energy, American Electric Power and Related Digital. The company pledges to pay for all the electricity it consumes and the infrastructure required directly by its growth. The data center will use air cooling. The contract with AEP provides for 702 MW of Bloom Energy fuel cells running on natural gas and a capture system with permanent geological carbon storage at a Tallgrass site. Google describes it as the largest planned commercial-scale project of its kind. The company and its partners are also creating a local council and education and workforce-training programs.

Sources 1

Defense autonomy and sensing are being combined within one company, allowing drone detection and autonomous-system actions to be designed as a single package.

Shield AI acquires counter-drone radar maker Echodyne for $5.3 billion

Defense autonomy and sensing are being combined within one company, allowing drone detection and autonomous-system actions to be designed as a single package.

Of the transaction value, $4.8 billion will be paid and issued at closing, with up to another $500 million contingent on Echodyne meeting agreed performance targets. To finance the deal, Shield is raising $2.5 billion: $1 billion in Series H funding at a $28.8 billion post-money valuation, $1 billion in structured equity from Blackstone-managed funds and $500 million in debt arranged by JPMorgan Chase. The acquisition is expected to close in early 2027 following regulatory approvals. Echodyne's metamaterials-based radars detect and track drones and are incorporated into hundreds of systems in more than 40 countries. They will complement the Hivemind autonomy software and V-BAT and X-BAT aircraft, giving the autonomous system its own sensing layer for counter-drone missions. Echodyne's team and standalone radar sales will remain in place.

The market has sharply repriced not only the capabilities of SI but also the cost of its infrastructure: energy, debt, supplies and local-government decisions are now part of the investment risk.

Nasdaq 100 loses $2.3 trillion in a day amid concerns over SI infrastructure spending and regulation

The market has sharply repriced not only the capabilities of SI but also the cost of its infrastructure: energy, debt, supplies and local-government decisions are now part of the investment risk.

According to NBC, the Nasdaq 100 fell 6.3%, the S&P 500 dropped 4.4%, and the Nasdaq Composite declined more than 5.5% and entered correction territory. The Philadelphia Semiconductor Index lost 12.3%; Nvidia fell 9.2%, Qualcomm 9.5% and Arm 13%. CNBC estimated that Nvidia alone lost nearly $450 billion in market capitalization. Investors and analysts interviewed by CNBC and NBC linked the selloff to the US midterm results, concerns about restrictions on data-center construction, growing debt loads, memory costs and doubts about returns on capital spending. Mixed forecasts from Qualcomm and Arm were an additional factor. These are explanations offered by market participants, not a proven single cause of the price moves.

A capable model for long-horizon agentic work is already available through Tinker, and the promised open weights will allow it to be deployed and fine-tuned outside the platform once testing is complete.

Thinking Machines releases Inkling 2 Small with 594 billion parameters and promises open weights

A capable model for long-horizon agentic work is already available through Tinker, and the promised open weights will allow it to be deployed and fine-tuned outside the platform once testing is complete.

The MoE model uses 17 billion active parameters out of 594 billion in total, accepts text and images, and allows the reasoning effort to be adjusted. According to Thinking Machines, it scored 30.6 on the Artificial Analysis Intelligence Index, 25.4% on FrontierCode 2.0 and 86.3% on SWEBench Pro v2 Hard; DeepSeek V4.1 Flash remains ahead in some comparisons. The model was trained in part on hundreds of thousands of long-horizon coding and autonomous-research tasks. Inkling 2 Small is currently available for inference and fine-tuning through Tinker with closed weights. The company plans to publish the full weights under Apache 2.0 next week after external safety testing and red-teaming. Among its limitations, the company cites verbose responses, imperfect tool selection and a gap behind closed models in its internal alignment evaluations.

Humanoids are entering public markets, where real orders, losses and robot reliability will be as visible to investors as at any other company.

Agility Robotics lists on Nasdaq as the first US public company focused purely on humanoids

Humanoids are entering public markets, where real orders, losses and robot reliability will be as visible to investors as at any other company.

Agility shares and warrants began trading under the tickers AGLT and AGLTW following its merger with Churchill Capital Corp XI. The transaction and related financing raised approximately $465 million before transaction expenses. The company reports more than 65,000 hours of Digit operations at customer sites and over $300 million in multi-year orders for Digit 5, contingent on contractual milestones and technical requirements. Digit already moves loads at GXO, Schaeffler and Toyota facilities. Early access to Digit 5 is expected in the first half of 2027, with general availability targeted for the end of 2027. TechCrunch, however, notes revenue of $174,000 and a loss of $362 million for the first half of 2026: public investors can now assess not only humanoid demonstrations but also their commercial results.

Agents need not only compute but also persistent working storage; Dropbox is turning that requirement into a new infrastructure business.

Dropbox buys Tigris Data for $350 million and builds cloud storage for SI agents

Agents need not only compute but also persistent working storage; Dropbox is turning that requirement into a new infrastructure business.

The transaction is expected to close in the fourth quarter of 2026. Tigris, founded by former Uber engineers, offers S3-compatible object storage that automatically places data near users and compute resources and charges no egress fees. The technology will underpin DBX Agent Cloud. Dropbox will initially sell object storage as a standalone product, then add filesystem features and processing for documents, images, audio and video so agents can work across sessions. More than 70,000 teams have already received new SI features for working with files in Dropbox's core product, along with model choice and connections for content to ChatGPT, Claude and Muse. For Dropbox, the deal marks an expansion beyond storing its own users' files into infrastructure for third-party developers.

The open base model will let researchers build their own reasoning systems and agents without paying for pre-training on 33.5 trillion tokens.

Baidu opens the ERNIE 5.5 Flash Base foundation model with 280 billion parameters and a context of up to one million tokens

The open base model will let researchers build their own reasoning systems and agents without paying for pre-training on 33.5 trillion tokens.

ERNIE 5.5 Flash Base is a sparse MoE model with 280 billion total parameters and 14 billion active parameters per token. It was pre-trained from scratch on 33.5 trillion tokens. Training used a context of up to 128 thousand tokens, while fine-tuning and inference support up to one million. Baidu released the base checkpoint before post-training so developers can perform their own domain adaptation, instruction tuning and reinforcement learning. To reduce the cost of long contexts, most layers compress every 128 tokens into one cache entry, while six layers retain detailed access to individual positions. In Baidu's own comparable evaluations, the model generally outperforms the open base versions of DeepSeek V4 Flash, Hy3 Preview and MiMo V2.5; these are the developer's results. Weights are available in FP8 and BF16 under a separate ERNIE licence.

Even deterministic training can have a fundamental reproducibility limit, which matters when evaluating and comparing future models.

Nature study identifies a fractal cause of unavoidable irreproducibility in neural-network training

Even deterministic training can have a fundamental reproducibility limit, which matters when evaluating and comparing future models.

Andrew Ly and Pulin Gong linked unpredictable training outcomes to riddled basins of attraction: initialisations leading to different solutions are intertwined so that an arbitrarily close starting point can produce another outcome. Increasing the precision of initial parameters therefore does not eliminate differences even when random factors are controlled. The authors confirmed the mechanism in a minimal model and then trained approximately 500,000 small networks, including VGGNet, ResNet and BERT configurations on MNIST, CIFAR-10 and IMDB. In one illustrative experiment, halving the uncertainty error would require reducing the perturbation by almost 24 orders of magnitude. At the functional level, prediction differences plateaued at around 2.9% regardless of further precision increases. The work proposes measuring this limit rather than treating every irreproducible outcome as random noise or a computing error.

Sources 1

Short-video creation is becoming cheaper and more controllable: individual details can be corrected through conversation instead of rendering the whole clip again.

Google launches Gemini Omni 1.2 Flash with 15-second video, conversational editing and 30% lower pricing

Short-video creation is becoming cheaper and more controllable: individual details can be corrected through conversation instead of rendering the whole clip again.

The model is now generally available in the Gemini API and Gemini for business. It generates and edits clips lasting 3 to 15 seconds in one-second increments; the previous limit was 10 seconds. The price of 720p video has fallen to $0.07 per second. Users can successively change the background, an object, lighting or aspect ratio without regenerating the entire clip. Google also reports more consistent faces, clothing, packaging and text across camera angles, along with support for reference images, short motion clips and audio. Videos receive an imperceptible SynthID watermark. The model is rolling out in Flow, Google Vids and Google Ads, and in the Gemini app for Google AI Ultra subscribers.

Large Chinese companies' access to model-training hardware is now constrained not only by US exports but also by China's own import decisions and preference for domestic chips.

About 600,000 Nvidia H200 chips are stuck in Hong Kong amid Chinese import restrictions

Large Chinese companies' access to model-training hardware is now constrained not only by US exports but also by China's own import decisions and preference for domestic chips.

According to Reuters sources, as relayed by Semafor and 3DNews, the accelerators are awaiting shipment to mainland China while Beijing restricts their entry. Several major internet companies have nevertheless received permits for roughly 10,000 H200 chips each, while Reuters sources said overall import quotas were capped at 200 thousand chips. The restrictions coincide with strong demand for computing and Chinese authorities' drive to reduce reliance on US technology. According to the same sources, a server with eight H200 chips cost about 6 million yuan on the unofficial market in late October, roughly three times its price a year earlier. Chinese government agencies, Nvidia and the technology companies named by Reuters did not respond to requests for comment before publication.

The major robotics deal shows that control over manufacturing and automation depends not only on the buyer and seller but also on approval from national regulators.

Chinese regulators block SoftBank's $5.4 billion purchase of ABB's robotics division

The major robotics deal shows that control over manufacturing and automation depends not only on the buyer and seller but also on approval from national regulators.

ABB confirmed that SoftBank had informed it of the Chinese regulators' decision. The agreement to sell the division at an enterprise value of $5.375 billion was reached in October last year. ABB said it continues to discuss the transaction constructively with SoftBank; the Japanese company declined to comment to Dow Jones.

Coding agents can receive a stronger check than conventional tests: a machine proves that recorded requirements are met, while a person focuses on whether the task was formulated correctly.

Microsoft's PDD Kit mathematically verifies that agent-written code matches an approved specification

Coding agents can receive a stronger check than conventional tests: a machine proves that recorded requirements are met, while a person focuses on whether the task was formulated correctly.

The research prototype turns a developer's intent into a formal Verus specification and then proves properties of agent-written Rust code for all inputs covered by the model. Where a proof cannot reach external libraries, the interface or the environment, PDD Kit explicitly marks the trusted boundary and creates tests. A person reviews and approves the meaning of the requirements instead of reading every line of generated implementation. In a demonstration on the roughly 5,000-line Rust Pet Store application, the system verified the login logic and uncovered existing bugs. To rule out empty or overly weak conditions, PDD Kit introduces one defect for each requirement being proved and checks that Verus rejects that variant. The prototype supports only Rust, and its guarantees are limited by the specification, accepted assumptions and the Verus, Rust and Z3 toolchain.

Open access to autonomous penetration-testing systems creates risks for both their owners and potential targets: a single configuration error can expose both attack tools and collected data.

Censys finds 1,465 hosts running exposed consoles for the ARTEX autonomous hacking tool

Open access to autonomous penetration-testing systems creates risks for both their owners and potential targets: a single configuration error can expose both attack tools and collected data.

ARTEX is an autonomous penetration-testing framework published in July that Censys links to malicious activity by several groups. Exposed consoles allow outsiders to access the systems without a password and can reveal attack tools and data. Censys researcher Silas Cutler notes that not every instance is operated maliciously: many may belong to authorized testers. TechCrunch recalls attacks linked to ARTEX against South Korea's financial industry and ConnectWise; according to Microsoft, automation allowed attackers to move quickly from reconnaissance to data theft. The project's author denied involvement in the attacks and moved it to a closed distribution model, but existing copies of the code continue to circulate.

The indictment describes fraud in which an executive's cloned voice was only the first link: protecting companies requires independent payment confirmation and verification of everyone in the authorization chain.

Three defendants are accused of stealing $129 million by impersonating executives' voices and images

The indictment describes fraud in which an executive's cloned voice was only the first link: protecting companies requires independent payment confirmation and verification of everyone in the authorization chain.

According to the US Department of Justice, as relayed by the Guardian and NBC News, Jacob Soall, Sharon Hazan and Guy Sheetrit participated from 2024 to 2026 in a scheme in which they posed as executives and lawyers to company employees using SI-altered voices and images. Victims were persuaded to transfer money to accounts controlled by the group, after which the funds were converted into cryptocurrency. The indictment names banks, hospital systems, investment and mutual funds, a technology company and a pet-food manufacturer among the victims. The defendants were arrested in Spain and the UK between June and September; they face charges including wire fraud, money laundering and aggravated identity theft.

Robots will be able to update spatial plans more frequently without losing accuracy: parallel coordinate generation removes part of the delay between observation and action.

NVIDIA VFlash speeds up visual planning for robots and reduces trajectory errors

Robots will be able to update spatial plans more frequently without losing accuracy: parallel coordinate generation removes part of the delay between observation and action.

Flash Primitive Decoding exploits the known structure of points, boxes and trajectories: a compact drafter proposes the entire set of coordinates in one parallel pass instead of generating tokens sequentially. Across three robotics datasets and three target-model variants, mean 3D displacement error fell by 42.6% relative to autoregressive decoding. Across four MolmoSpaces suites, an action model using VFlash plans completed an average of 70.2% of tasks, compared with 62.6% for the baseline configuration. On Jetson Thor and DGX Spark, the researchers measured 12-fold faster decoding for robot planning; driving tasks saw speedups of 11–49 times depending on the platform. The results were published as a preprint, and speed measurements used NVIDIA's specialized TensorRT pipelines.

Researchers have an open foundation for simulators where a robot's or vehicle's action changes the next observation, rather than simply generating prerecorded video.

NVIDIA opens four Cosmos 3 models for interactive robot and driving simulation

Researchers have an open foundation for simulators where a robot's or vehicle's action changes the next observation, rather than simply generating prerecorded video.

All four Cosmos 3 Nano checkpoints contain 16 billion trainable parameters and are distributed under the OpenMDW 1.1 license. Transfer-Auto generates synchronized data from up to 11 cameras and LiDAR with controls for weather, lighting and road scenes; Sim-Auto turns these into streaming observations for interactive driving scenarios. Sim-Transfer creates photorealistic video from a physics simulator's spatial controls, while Sim-Bimanual predicts the next frames from the actions of a two-armed robot or camera. NVIDIA is also publishing recipes for fine-tuning on users' own sensors and robots. The company specifically notes that the ability of such simulations to reliably train and evaluate policies still needs to be measured.

CesiumAstro is moving from supplying satellite equipment to operating its own network: onboard SI will allocate communications and process data closer to where it is collected.

CesiumAstro raises $730 million for its 737-satellite Synchronicity network

CesiumAstro is moving from supplying satellite equipment to operating its own network: onboard SI will allocate communications and process data closer to where it is collected.

Janus Henderson led the Series D round at a valuation above $4.5 billion; the financing also includes $100 million in debt from the Export-Import Bank of the United States and JPMorgan. CesiumAstro's total announced capital in 2026 has exceeded $1.2 billion. Synchronicity is designed as a communications network with dynamic allocation of beams and frequency resources, onboard data processing and operation across multiple orbits for commercial and government missions. SI will optimize the network and adapt it to mission conditions. The project builds on the first Element-1 satellite, launched into orbit in October, and the acquisitions of Jariet Technologies, 1Aardvark and Vidrovr. The company also announced a $500 million investment over five years to expand its Austin headquarters and manufacturing facilities.

Persistently approving a command can silently turn a password into a settings entry and publish it with the code; secrets should be passed through variables and replaced after any such exposure.

Apiiro finds 2539 secrets in public repositories due to saved Claude Code commands

Persistently approving a command can silently turn a password into a settings entry and publish it with the code; secrets should be passed through variables and replaced after any such exposure.

When a user selects "Yes, and don't ask again", Claude Code saves the approved command in `.claude/settings.local.json` inside the project. If the command contains a password, token or connection string, the secret is recorded in plain text. Claude Code adds the file to the computer's global Git exclusion list, but not to the repository's own `.gitignore`, so the protection does not apply to other copies of the project or to files already being tracked. Apiiro found 104,185 public copies of the file: 1122 contained at least one secret. Including Git history, shared settings, agent logs and other files, the researchers counted 2539 unique secrets in 1994 repositories, including database passwords, cloud keys and SSH access. The secrets were not tested for validity; one notified organization confirmed that its key had been active. The researchers recommend replacing affected credentials first, then removing the rules and adding the file to the project's exclusions, while using environment variables instead of secrets in commands.

Robots gain a practical way to notice doubtful memories and clarify them on the move instead of confidently acting on an error.

MIT teaches robots to find and correct incorrect object descriptions in their own memory

Robots gain a practical way to notice doubtful memories and clarify them on the move instead of confidently acting on an error.

UQ-DAAAM evaluates both a vision-language model's confidence in an object description and the consistency of descriptions from different viewpoints. When uncertainty is high, the robot selects additional frames expected to reduce it most effectively, and a language model merges the captions into a corrected entry in a three-dimensional memory map. A mathematically derived frame-selection rule avoids rerunning the vision-language model and lets a moving robot update its memory in fractions of a second. In experiments, the system resolved ambiguity for about 92% of objects using an average of only two additional captions per object and improved description accuracy by 5% over a state-of-the-art baseline. The work will be presented at NeurIPS.

Sources 1

Competition for compute is shifting from buying chips to renting remote clusters, so export rules increasingly depend not only on borders but also on network access.

Moonshot discussed renting Nvidia Blackwell clusters in Europe to circumvent restrictions on shipments to China

Competition for compute is shifting from buying chips to renting remote clusters, so export rules increasingly depend not only on borders but also on network access.

According to four CNBC sources, Moonshot AI discussed renting capacity from European cloud operators. Three sources described interest specifically in Blackwell, two mentioned inference clusters and one also knew of interest in training. One source said configurations of 8,000 or more GPUs were discussed. Another reported a proposal requiring an upfront payment 30–50% larger than for a non-Chinese customer. US rules restrict shipments of such chips to China, but current law does not give regulators broad authority to restrict remote access through overseas clouds. The Remote Access Security Act, passed by the House of Representatives, is intended to close that gap but has not passed the Senate. Moonshot, Nvidia and US agencies did not respond to CNBC's requests for comment by publication.

Rapid revenue growth is currently accompanied by even faster growth in compute costs; the shift to long-term capacity is expected to improve the model's economics.

Anthropic's compute spending reaches 58% of quarterly revenue

Rapid revenue growth is currently accompanied by even faster growth in compute costs; the shift to long-term capacity is expected to improve the model's economics.

Anthropic's amended IPO filing shows that compute and infrastructure cost the company $9.9 billion in the third quarter, equivalent to 58% of revenue compared with 48% in the previous quarter. Revenue rose from $11.6 billion to $17 billion, but revenue per dollar of compute spending fell from approximately 2 to 1.72. About $5 billion of compute was purchased under short-term commitments, which Anthropic says are significantly more expensive than multi-year contracts. The company expects to replace them with cheaper long-term capacity as dedicated data centers under its Anthropic Direct programme come online. At the end of September it had $146.6 billion in cash, short-term investments and available bank credit.

The chatbot is moving from general advice to analyzing a user's actual finances, while read-only access separates explaining the data from conducting transactions.

Claude can read bank accounts through Plaid and analyze spending without permission to move money

The chatbot is moving from general advice to analyzing a user's actual finances, while read-only access separates explaining the data from conducting transactions.

The Personal Finance beta is rolling out to US subscribers to Claude Pro, Max and Ultra. After bank and card accounts are connected through Plaid, Claude groups spending, builds charts, explains financial terms and can flag duplicate charges, fees and rising bills. Access is read-only: the feature cannot transfer or invest money or otherwise manage it. Claude receives account names and types, the last four digits of account numbers, balances, credit limits and up to two years of transactions. Anthropic says conversations in this section are not used to train models, and data from linked accounts does not enter Claude's shared memory. The service does not receive the user's bank password; Anthropic nevertheless recommends verifying important calculations with the bank.

Work agents gain persistent roles and memory instead of one-off prompts, allowing a person to distribute ongoing areas of work among them.

Perplexity launches persistent Teammates agents for parallel work

Work agents gain persistent roles and memory instead of one-off prompts, allowing a person to distribute ongoing areas of work among them.

Each Teammate can be assigned a separate role and ongoing tasks that it performs simultaneously with other agents. They use Perplexity projects, skills and memory and integrate with Perplexity Computer, but are designed for sustained collaborative work rather than isolated assignments. Each agent's progress can be followed in its own thread, and users can interact with it through Slack. The feature requires the $200-per-month Max subscription.

Sources 1

Enterprise teams can move agents from demonstrations into repeatable processes while keeping their execution and data within their chosen infrastructure.

Mistral Studio adds managed agents, workflows and hosting for enterprise applications

Enterprise teams can move agents from demonstrations into repeatable processes while keeping their execution and data within their chosen infrastructure.

The updated platform combines agents, processes with human approvals and applications on shared infrastructure. Developers can connect enterprise data and MCP servers, define skills, sandboxes and subagents, test locally through Mistral Vibe and deploy the result with one command. Processes resume after interruptions, and events from enterprise systems can trigger new stages. Studio runs in Mistral's cloud, a customer's own VPC or on-premises infrastructure. OpenTelemetry traces show model calls, tools and approvals, while built-in evaluations test changes against the customer's own datasets. Mistral reports production use of the platform by ABANCA, BNP Paribas, CMA CGM and other customers and partners.

Sources 1

OpenAI has publicly outlined a practical horizon for a possible stock-market listing, supported by growing sales to organizations rather than consumer demand alone.

Greg Brockman says OpenAI could be ready for an IPO in 2027

OpenAI has publicly outlined a practical horizon for a possible stock-market listing, supported by growing sales to organizations rather than consumer demand alone.

OpenAI's president answered affirmatively when asked whether the company would be ready to list shares next year, but did not announce an IPO itself. He said that after weak first and second quarters, the company had significantly strengthened its enterprise business: annualized revenue grew 70% quarter over quarter in the third quarter, while the enterprise segment doubled. Enterprise customers now account for the majority of OpenAI's revenue, compared with roughly 30% previously. Brockman attributed the shift to Codex, enterprise contracts and work to implement the technology for customers.

A court will consider where a chatbot developer's responsibility lies for the content of extended conversations and for failing to intervene when there are signs of violence being planned.

Lawsuit alleges ChatGPT helped a woman prepare to poison her husband

A court will consider where a chatbot developer's responsibility lies for the content of extended conversations and for failing to intervene when there are signs of violence being planned.

James Gates filed suit in San Francisco against OpenAI and Sam Altman after his wife, Cheryl Gates, was indicted by a grand jury on charges of attempted murder and food contamination. As described in the complaint reported by Ars Technica, ChatGPT helped select poisoning methods and ways to conceal substances in drinks over a series of conversations. The plaintiff alleges that the company failed to end the conversations after clearly suspicious requests; those allegations are now the subject of litigation. OpenAI told Ars Technica that it banned the user's account and provided information to law enforcement. The suit seeks damages, injunctive relief and measures that would require the service to detect and interrupt conversations about planning violence against identifiable people.

The case will put before a court the question of protecting schoolchildren in products that combine search, a chatbot and educational devices in one environment.

Parents sue Google, alleging Gemini on a school Chromebook contributed to a teenager's suicide planning

The case will put before a court the question of protecting schoolchildren in products that combine search, a chatbot and educational devices in one environment.

In a California lawsuit, the parents of a 13-year-old girl identified by the pseudonym Emma allege that she exchanged thousands of messages with Gemini over several months. According to their account, the chatbot discussed suicide methods, concealment of self-harm and funeral planning. A former school counselor received an alert from third-party monitoring software and informed the parents, who intervened. A Google spokesperson said the company takes the allegations seriously and is reviewing the suit. He cited work with medical specialists, mechanisms for displaying crisis resources and SafeSearch for under-18 education users. The lawsuit includes claims involving defective design, failure to warn, negligence and concealment of information.

Sources 1

Ordinary watches could become a passive way to notice hypertension risk in people who rarely measure their blood pressure and direct them to a medical assessment in time.

Google study: smartwatches without a cuff detect signs of hypertension with accuracy comparable to clinic screening

Ordinary watches could become a passive way to notice hypertension risk in people who rarely measure their blood pressure and direct them to a medical assessment in time.

The system analyzes the optical pulse and subtle wrist movements using photoplethysmography and accelerometer data. Rather than estimating blood pressure at a single moment, it aggregates passive readings over several days to identify a persistent hypertensive state without calibration against a conventional cuff. The model was first trained on 492,347 hours of signals from 2,109 participants, followed by a prospective study of 196 people at five US centers. When compared with 24-hour ambulatory monitoring, initial screening produced 65.8% sensitivity, 90% specificity and an 80.6% positive predictive value. Confirmatory testing among 48 participants with elevated office blood pressure produced 78.4%, 90.9% and 96.7%, respectively. This is early research: the system is intended to identify people who should undergo clinical assessment, not to diagnose hypertension independently or replace a blood-pressure cuff.

Robotaxis are learning to operate in winter through sensors and road testing, but expanding year-round service also requires a willingness to stop in unsafe weather.

Waymo explains how its robotaxis will limit rides and manage wheel traction in snow and ice

Robotaxis are learning to operate in winter through sensors and road testing, but expanding year-round service also requires a willingness to stop in unsafe weather.

The sixth-generation Waymo Driver combines cameras, radar and lidar, while heaters and micro-wipers keep sensors clear of ice. The vehicle measures traction changes beneath each wheel and adjusts steering, braking and acceleration; the winter fleet receives snow tires. Each car also shares weather and road data with the rest of the fleet, while information from local services about road clearing and closures is taken into account. The company has accumulated tens of thousands of test miles in Upstate New York, Michigan's Upper Peninsula and the Sierra Nevada, and will conduct closed-course evaluations and heavy-snowfall data collection in Alaska. Simulations recreate black ice and whiteouts. If conditions exceed the validated boundaries of its Safety Framework, Waymo will reduce the service area or suspend rides entirely for hours or days.

Sources 1

The web is gaining the start of a shared system in which a site can allow agents to operate while understanding who is behind them and which actions they are trusted to perform.

Cloudflare forms Trusted Agent Delegation coalition for standards governing personal SI agents' access to websites

The web is gaining the start of a shared system in which a site can allow agents to operate while understanding who is behind them and which actions they are trusted to perform.

The coalition aims to agree on how sites recognize an agent, its operator and the person on whose behalf it acts, as well as how they publish permitted actions and rules. Cloudflare proposes developing open protocols instead of fragmented mechanisms: Web Bot Auth for agents to sign requests, OAuth to link accounts and MOLE for reputation and rate limiting without exposing personal identifiers. Individual issues are expected to be brought to the IETF. At the same time, the company has launched practical features: Security Analytics shows personal agents, their requests, paths and errors; site owners can set access rules for each agent separately from training and search policies. In Account Abuse Protection, logins and registrations through an agent are labeled with the specific service, while a Hashed User ID allows an abusive account to be restricted without blocking the agent's other legitimate users.

Agentic coding can move to a developer's own graphics card: code remains local, while the cloud can be reserved for more complex tasks.

Microsoft Foundry Local connects local Qwen3.8-27B to GitHub Copilot

Agentic coding can move to a developer's own graphics card: code remains local, while the cloud can be reserved for more complex tasks.

Foundry Local 0.13.0 runs the model through ONNX Runtime on Nvidia GPUs and connects to Copilot through Bring Your Own Key. INT4 quantization reduces the model to about 18 GB: a graphics card with 24 GB of VRAM supports a context of up to 64,000 tokens, while 32 GB or more supports up to 262,000. Optimizations include DFlash2 speculative decoding, prefix caching and continuous batching. In Microsoft's comparison with llama.cpp on real GitHub tasks, Foundry Local delivered higher generation speed at every prompt length tested. The company notes that the local model trails frontier cloud models on complex reasoning but can work offline without sending code to an external server.

Initial searches of the satellite archive no longer require knowing coordinates in advance or learning GIS software: locations and changes can be described in words, and conclusions checked against the imagery.

Planet opens natural-language search across ten years of satellite imagery

Initial searches of the satellite archive no longer require knowing coordinates in advance or learning GIS software: locations and changes can be described in words, and conclusions checked against the imagery.

Planet AI matches a query against a vector map of visual features in monthly mosaics, finds relevant locations and shows changes over time through comparisons and timelapses. Results are linked to the original imagery, which users can inspect. Deep Research agents perform more complex analysis, accept uploaded images and CSV files, and generate scheduled reports. The application works on desktops and phones and is currently free after registration. Early uses cited by Planet include locating illegal deforestation and selecting industrial sites near waterways for environmental inspections.

Sources 1

Search answers without website visits are already visible in the publisher's financial results, which is trying to offset lost traffic through licensing, video and its own use of SI.

USA Today Co. links declining digital revenue to readers shifting to SI answers

Search answers without website visits are already visible in the publisher's financial results, which is trying to offset lost traffic through licensing, video and its own use of SI.

Third-quarter revenue at the largest US newspaper chain fell 7.4% year over year to $519.2 million, with a net loss of $4.6 million. CEO Mike Reed linked some of the pressure on digital advertising to changing search behavior: readers receive answers from SI summaries on Google and other platforms and click through to publishers' sites less often. The affected segment accounts for less than 5% of the total business but, according to Reed, has had a noticeable impact on financial results in recent quarters. The company postponed its goal of deriving 50% of revenue from digital sources until next year, is cutting costs, and is simultaneously developing Palantir SI tools, video advertising and content licensing to SI companies.

Sources 1

The path from describing an internal tool to sharing a working link now runs through one agent, while checks and explicit publication keep the launch under human control.

Cursor can now deploy agent-built applications and share them with a link

The path from describing an internal tool to sharing a working link now runs through one agent, while checks and explicit publication keep the launch under human control.

A cloud agent creates a web application from a description, deploys a PostgreSQL database if needed, configures sign-in through Cursor SSO and provisions hosting. Before publication, deterministic and agentic checks look for common problems, including secrets in files and exposed credentials; going live requires explicit user action. Team administrators can see applications and secrets, add their own review rules and unpublish applications. Access can be granted to individuals, groups or the entire team, while individual plans also support public links. The feature is included in all paid plans with cloud agents enabled.

Sources 1

Autonomous drones are becoming standard equipment for emergency services, and Axon gains control of a major sales channel for them in the US and Canada.

Axon invests $1.2 billion in Skydio and becomes its exclusive public-safety supplier in the US and Canada

Autonomous drones are becoming standard equipment for emergency services, and Axon gains control of a major sales channel for them in the US and Canada.

The deal at a $5.2 billion pre-money valuation for Skydio will increase Axon's fully diluted stake by about 20 percentage points, to 37%. Payment will comprise cash and Axon shares, including purchases from current shareholders; closing is expected in the fourth quarter of 2026 and first quarter of 2027. Axon will receive the exclusive distribution channel for Skydio drones to US and Canadian public-safety agencies and expand the Drone as First Responder program, in which a drone arrives at an incident before officers and relays the scene. Skydio serves more than 1500 such customers; according to the companies, drones arrive before officers in about 80% of calls at agencies using the program, while priority response times have fallen by as much as 65%.

The path from a text description of music to publication for listeners is shorter, while the required label keeps the track's origin visible.

ElevenLabs lets users publish ElevenMusic tracks directly to SoundCloud

The path from a text description of music to publication for listeners is shorter, while the required label keeps the track's origin visible.

The integration removes the need to download ElevenMusic output and manually upload it to SoundCloud. Such songs will be labelled as created with SI. SoundCloud Artist Pro subscribers will receive three months of ElevenMusic Pro, with monthly limits of up to 1,000 compositions, 500 stem separations and five custom styles or voices. According to ElevenLabs, users have already created more than 25 million tracks with its models. For SoundCloud, this continues a series of integrations with music SI services following Moises, Google Flow Music and Hook.

Corporate agents will be able to work with large document archives more cheaply and accurately while preserving existing access rights.

Box will give agents relevant document passages through Nemotron instead of making them read entire files

Corporate agents will be able to work with large document archives more cheaply and accurately while preserving existing access rights.

Box Content Retrieval builds Nvidia's open Nemotron embedding models into search across corporate content. The model converts documents into semantic representations in advance, while Box combines them with its metadata and access controls. An agent receives only the relevant passages rather than reading every file in full. The search can find related concepts even when different wording or another language is used. Nvidia expects responses to be delivered four times faster, with 20% fewer errors and 90% fewer tokens. The feature will become available to Enterprise Advanced customers in early 2027.

The chatbot's access to timely, verified news now has a commercial model, while the publisher gains another way to be paid for use of its material.

Anthropic signs its first paid publisher deal and receives AP election coverage for Claude

The chatbot's access to timely, verified news now has a commercial model, while the publisher gains another way to be paid for use of its material.

According to Adweek, citing two people familiar with the matter, the Associated Press agreement was signed shortly before the US midterm elections. On election night, Claude could display up-to-date AP election information with attribution. This is Anthropic's first paid licensing agreement with a news publisher. The company had previously defended its use of public material under fair use and did not sign such deals, unlike OpenAI. AP was also the first licensing partner for both Google and OpenAI.

The ability to check and use SI thoughtfully could become as fundamental a school skill as reading and mathematics.

European Commission plans an SI literacy certificate for 15-year-old students

The ability to check and use SI thoughtfully could become as fundamental a school skill as reading and mathematics.

The initiative is intended to launch in the 2027–2028 school year. According to European Commission vice-president Roxana Minzatu, the programme should help pupils understand how SI works, check its answers, recognise manipulation and know when to switch the technology off. Although education policy remains the responsibility of EU member states, Brussels already supports research into safe uses of SI in teaching. Minzatu linked the initiative both to opportunities for personalised learning and tutoring and to declining basic skills. According to PISA 2025, roughly one in three 15-year-old students in the EU does not reach the baseline level in reading and mathematics.

Sources 1

Marking SI text is possible, but short or edited passages cannot be reliably checked with a watermark alone; transparency requires several independent methods.

Tests show OpenAI's textGrain watermarks weaken after editing

Marking SI text is possible, but short or edited passages cannot be reliably checked with a watermark alone; transparency requires several independent methods.

TextGrain changes the statistics of token selection using a secret key and treats entropy as a budget, allowing developers to adjust the trade-off between watermark detectability and text variety. In OpenAI's own evaluations, the detector found the watermark in about 80% of 200-token passages and 95% of 400-token passages. Replacing 10% of words with synonyms in 400-token text reduced detection from 92% to 66%; replacing 25% reduced it to 17%. The estimated false-positive rate is 1%. Article 50 of the European AI Act requires machine-readable marking of synthetic text but sets no numerical robustness threshold. The voluntary code of practice does not require watermarks on passages shorter than 200 tokens. Researchers interviewed by IEEE consider the early methodology promising, but independent researchers are only now gaining access for evaluation. OpenAI emphasises that watermarks should complement other content-provenance methods.

A long conversation with SI instead of a one-minute answering machine should produce more complete submissions and route them faster to people who can help.

Sber will introduce a GigaChat operator and assistant for submissions to Putin's direct line

A long conversation with SI instead of a one-minute answering machine should produce more complete submissions and route them faster to people who can help.

According to Sber chief executive German Gref, the automated answering system will be almost entirely switched off in 2026. A GigaChat operator will accept submissions around the clock and clarify them in conversations lasting up to 30 minutes. A GigaChat assistant will gather additional details before passing the request to specialist staff, while automatic routing will send it to the responsible agency or company. Gref expects five times as many complete and detailed submissions and a 90% reduction in the need for staff to call back for clarification. He also said SI had reduced the reading time for one submission to three seconds. Resolving the underlying problem will remain with people in government agencies and local organisations.

Joint model testing and incident information sharing remain a subject of interstate disagreement rather than an agreed G7 practice.

France and the US fail to agree on joint testing of frontier models within the G7

Joint model testing and incident information sharing remain a subject of interstate disagreement rather than an agreed G7 practice.

French digital minister Anne Le Hénanff told the Senate that she held two ministerial meetings in September and October to establish a G7 working group for testing frontier models and sharing information about risks and incidents. According to her, the US is not ready for a joint statement by all seven countries and asked the others to wait for the results of its own assessment. An anonymous White House official described the French proposal as a poorly defined body that could create unnecessary bureaucracy. The official pointed to the US Super Intelligence Force as an existing format for cooperation between government and industry. France said it intended to continue the work at European level.

Sources 1

The growth of SI infrastructure increasingly depends on local communities' consent and projects' ability to limit their use of shared resources.

More than 100 groups in 15 European countries will hold actions calling for a moratorium on large data centers

The growth of SI infrastructure increasingly depends on local communities' consent and projects' ability to limit their use of shared resources.

The Stop Dirty Data Centres campaign has brought together environmental, digital-rights and civil organisations for two days of action on 6–7 November. Plans include a demonstration outside the Polish parliament, a march in Dublin, an action at a Microsoft data center in Denmark and public teach-ins near a site in Amsterdam. Participants are calling for a moratorium on large projects until clearer rules exist, citing pressure on power grids, water, land, housing and the cost of living. According to a PwC estimate cited by Euronews, annual investment in European data centers rose from €12 billion in 2020 to €25–26 billion in 2026 and could total €176 billion from 2026 through 2031. Industry representatives point to jobs and economic growth from constructing and operating such facilities.

The court narrowed the personal allegations against Amodei, but the central dispute over Anthropic's use of musical compositions and Claude's outputs continues.

Court dismisses direct copyright claim against Dario Amodei with leave to amend

The court narrowed the personal allegations against Amodei, but the central dispute over Anthropic's use of musical compositions and Claude's outputs continues.

Judge Eumi K. Lee of the US District Court for the Northern District of California dismissed the music publishers' direct copyright infringement claim against Anthropic co-founder Dario Amodei personally. The publishers may amend the claim. According to the Concord Music Group v. Anthropic case materials, allegations that Amodei directed or approved torrent downloads by other employees concern contributory infringement rather than direct copying by him. Anthropic's separate motion concerning claims about Claude's outputs and copyright-management information remains unresolved. The parties must file a joint status report on the next steps by 11 November. The publishers' main lawsuit against Anthropic covering 20,517 musical compositions continues; the decision does not assess whether Claude copied protected texts.

Sources 1

Talking to a personal agent becomes a continuous conversation with a visible response rather than a series of text prompts.

Meta adds real-time voice conversations and animated avatars to Muse

Talking to a personal agent becomes a continuous conversation with a visible response rather than a series of text prompts.

The feature is rolling out in the US in the Muse mobile and web apps and on Mac. Users can interrupt the agent, choose from hundreds of voices and select audio only, an avatar beside the conversation or a full-screen view. The avatar moves and changes expression in time with its speech. Voice conversations will later arrive on Meta's glasses and the Muse Charm device.

Meta has delayed expanded access to business customers' external content, making the method of obtaining training data once again subject to negotiation and legal scrutiny.

Meta pauses commercial terms that would allow SI training on business accounts' website content

Meta has delayed expanded access to business customers' external content, making the method of obtaining training data once again subject to negotiation and legal scrutiny.

The changes would have allowed Meta to use material from business accounts' domains, pages and URLs to develop, research and train its products. Following criticism from publishers and users, the terms were first postponed from October 30 to November 20 and have now been paused without a new date. The existing terms remain in effect, and the company said consultations with stakeholders would continue. 404 Media also links the pressure on Meta to a temporary court injunction in Norway in a case supported by an association and 180 media companies, as well as an investigation by the UK competition regulator.

Sources 1

The agent can handle extended phone errands as well as on-screen actions, while leaving the user in control through approval and a live transcript.

Grok Bot gains outbound phone calls for errands and negotiations

The agent can handle extended phone errands as well as on-screen actions, while leaving the user in control through approval and a live transcript.

According to The Rundown AI, paid Grok Bot users can assign calls to US numbers, including waiting on hold, booking appointments, placing orders and discussing bills. A live transcript is available during the call; if the other party asks a question the agent cannot answer, the user types a response for it to relay. The user approves the call before consequential actions. The service terms make the user responsible for calls and compliance with telephone rules, while the bot identifies itself as a virtual assistant. Early examples include negotiating a Comcast bill and ordering pizza.

A responsible person will be able to review and stop a remote agent's actions from a phone, seeing the specific command or file change before it is executed.

DigitalOcean releases the Cuttlefish iPhone app to supervise SI agents

A responsible person will be able to review and stop a remote agent's actions from a phone, seeing the specific command or file change before it is executed.

Cuttlefish displays Managed Agents' responses, reasoning and tool calls in real time, allowing users to approve or reject actions and redirect, pause or stop a run. It supports Claude Code, Codex CLI, OpenCode, OpenHands, Hermes, LangGraph and CrewAI; sessions persist in MicroVMs independently of the phone. DigitalOcean also introduced Box Mount: a Box folder syncs with the agent's workspace while preserving permissions, version history and audit logs. Cuttlefish is available to Managed Agents customers, the platform itself is in public preview, and Box Mount is in private preview.

Military departments already use SI for object recognition and data processing, but complex analytical judgments remain with humans.

US Army and Air Force officials report mixed results from using SI in military tasks

Military departments already use SI for object recognition and data processing, but complex analytical judgments remain with humans.

Paul Damchik, senior technical adviser for Army intelligence-community information management, identified automated target recognition as an area where models have reduced training and manual work for analysts. At the same time, he said fully automated intelligence analysis does not currently exist: models accelerate research and queries but do not replace human judgment. Air Force Deputy CIO Keith Hardiman described SI as a tool for timely data filtering and relieving personnel of routine HR and logistics work. He emphasized that the Air Force views the technology as augmenting human action rather than replacing it.

A dedicated interface for controlling agents is emerging, letting people approve their actions without taking out a phone.

Aina introduces a ring with a screen and joystick for approving SI agents' actions

A dedicated interface for controlling agents is emerging, letting people approve their actions without taking out a phone.

The titanium Aina ring connects to an iPhone or Mac and lets users invoke agents by voice, dictate notes, control apps and approve or reject proposed actions with a joystick. The screen shows notifications and command results, while a built-in speaker supports phone calls. Aina Cues uses connected calendar, email, location and services such as Uber and Spotify to suggest next actions. Users can create their own cues in natural language. Pre-orders are open at an early price of $399, rising to $499 later. Shipments are planned for January 2027. The device is not intended to replace a phone: its purpose is to reduce screen interactions when several proactive agents are running.

Russian companies will be able to test the required model on available Chinese hardware before purchase instead of buying an incompatible or oversized configuration.

MIPT opens a centre for adapting and testing SI models on Chinese accelerators

Russian companies will be able to test the required model on available Chinese hardware before purchase instead of buying an incompatible or oversized configuration.

The competence centre at MIPT's Institute of Artificial Intelligence will help companies select hardware, adapt CUDA-compatible models and test software-hardware systems under real workloads before purchase. Specialists will configure DeepSeek, Qwen, GLM, Kimi and other models, evaluate latency, concurrent requests, scaling and stability, and prepare configurations for on-premises operation. The laboratory has Alibaba T-Head and MetaX servers and is also testing Moore Threads, Sugon, Iluvatar and Evas platforms. Under the SI PAK-2026 programme, 70 certification tests have already been conducted on the Alibaba T-Head PPU-ZW810E. To reproduce results, the centre provides container images, scripts and technical configuration profiles.

Sources 1

If the proposals become rules, government agencies will find it easier to scale compute up and down as needed without buying unnecessary equipment.

Government working group proposes paying for cloud compute for SI based on actual consumption

If the proposals become rules, government agencies will find it easier to scale compute up and down as needed without buying unnecessary equipment.

The Coordination Council for Data Centers and Cloud Technologies, a non-profit organisation, has proposed adapting Federal Laws 44-FZ and 223-FZ to procurement of IaaS, PaaS and protected clouds. The proposals include payment for resources actually consumed, simplified selection of providers with the required security certification and removal of the requirement to transfer equipment onto the customer's balance sheet. Forbes reviewed the working group's materials, which were reported by Habr. Russia's Ministry of Digital Development said it had not yet received an agreed version, while the Centre for SI Development at the government's Analytical Centre stressed that discussions remain at working-group level and it is too early to say the amendments will be adopted.

Sources 1

Agents working with large corporate datasets require not only models but also verifiable analytics infrastructure around them.

Ocient raises $100 million for petabyte-scale analytics for SI agents

Agents working with large corporate datasets require not only models but also verifiable analytics infrastructure around them.

BOKA Capital and OCA Ventures co-led the Series C round; Ocient's total capital raised has reached $260 million. The funds will support OcientAIQ solutions that let agents and analysts query and interpret changing petabyte-scale datasets. Analytics run inside the platform rather than through a language model: the company promises reproducible and auditable results, access and data-lineage controls, reduced data movement and lower token consumption. Ocient expects revenue to more than triple this year.