Every story, at length

The new model sits at the top of OpenAI’s Daybreak programme, which gives cybersecurity defenders access to its most capable systems. ChatGPT Cyber will be available through Daybreak Red, a tier for offensive researchers hunting and developing vulnerabilities. OpenAI said the model had already scanned the Linux kernel, the core software behind many computer systems, and prepared patches after finding 948 potential issues. It also found 81 “novel” vulnerabilities in the Firefox web browser.

The announcement follows a striking demonstration by Max Justicz, chief technology officer and co-founder of Charge Robotics. Using GPT 5.6 Sol through Daybreak Blue, he asked ChatGPT to examine code where he had previously found a bug and look for problems “as bad or worse”. Two hours later, it returned a severe vulnerability that could have let hackers take control of servers running Ubuntu, a widely used Linux-based operating system. Justicz reported it to the Debian Project, which issued a patch; there is no known evidence that it was exploited. He spent about $5,000 on subscriptions and tokens to find several exploits.

For companies and open-source maintainers, the practical shift is that useful security research may no longer require a large specialist team. OpenAI says it wants to expand Daybreak to organisations of every size, independent researchers and open-source maintainers, and has announced a $1 billion initiative supporting participants who protect critical infrastructure such as electrical grids and water systems. Yet only “thousands” of cybersecurity professionals are in the programme, compared with a far larger pool of potential attackers.

The genuinely new element is broader, structured access for defenders, rather than proof that AI can find bugs. The limits are also clear: Justicz said Daybreak Blue appeared to have been weakened recently, while joining required only a photo of his driver’s licence. OpenAI researcher Eric Wallace said the company was devoting significant frontier computing power to finding and patching open-source flaws free of charge. But the Linux bug hunt cost $1.2 million, against roughly $6 million in token spending on a possible solution to a famous fluid-dynamics problem.

The $1.6 billion startup announced the work on Wednesday. Stacks is an open-source reference started in 2005 by Columbia mathematician Johan de Jong, with contributions covering college-level algebra through advanced algebra, geometry and number theory. Columbia mathematician Peter Woit has compared it with the Bourbaki books, which rebuilt mathematics from its foundations in the 20th century.

Formal verification means rewriting proofs in a programming language called Lean, then having a proof assistant check the code line by line. Checking individual theorems was only part of Axiom’s task: the results build on one another, so the team had to preserve their connections throughout. Engineers already use the same approach to confirm chip designs, while researchers in France have used it to catch bugs in software calculating family benefits.

The most striking detail is how little direction the agents needed. Axiom imposed a loose hierarchy, then let them find tasks and work in parallel. The company also trained its own models for mathematics, initially using an in-house Lean expert to teach them a human sense of taste; CEO Carina Hong says they later developed it themselves. Hong, 25, left a Stanford mathematics PhD programme and founded Axiom in March 2025. The company raised a $200 million Series A led by Menlo Ventures in March.

The work is new in scale and coverage, rather than in the underlying checking method. Axiom technical staff member Ashvin Swaminathan says algebraic geometry has been a major gap in formalised mathematics, leaving researchers little verified material on which to build. Axiom’s commercial bet is broader: as AI agents write more software, reusable checked results could help them avoid compounding errors. Whether that mathematical foundation transfers to dependable business software remains a bet, but Hong has openly challenged coding companies to attempt formalisation at this scale.

Anthropic researchers analysed 207,049 Claude and Claude Code sessions from consumer plans between August 2 and September 30, 2026, using privacy-preserving classifiers that retained only category labels such as field and task, not research content. Data analysis accounted for 33.7% of Claude sessions and 48.1% of Claude Code sessions, against 27.6% of researcher time. Writing and communication made up 32.0% of Claude sessions, double its 16.0% share of research time. Ideation and hypothesis generation drew just 4.2% of Claude sessions and 2.1% on Claude Code, below the 7.1% of time researchers spend on it.

The mismatch is striking because an accompanying survey of 528 scientists, mostly active AI adopters in life and health sciences, ranked idea generation first and data gathering second for producing important research results. Data gathering took 28.6% of researcher time but only about 12% of sessions, partly because clinical and biochemical collection involves physical work. The authors caution that ideation may be undercounted because it is woven into other tasks, but scientists may also prefer to generate ideas themselves or doubt AI's ability to choose questions and design studies.

Usage closely followed the report's measure of theoretical AI capability. Tasks in the top 20% received about twice as much Claude use and four times as much Claude Code use as researcher time would predict; the bottom 20%, including most physical work, received very little. Computer science supplied 32.7% of Claude Code sessions despite representing 4.3% of the world's researchers. Artificial intelligence and machine learning accounted for 15.2% of Claude Code sessions from just 1.2% of researchers.

The genuinely new evidence is this task-level map of actual use, rather than another demonstration of an AI breakthrough. Anthropic's scenarios suggest widespread adoption of existing capabilities could double output in some fields. Full automation of non-physical work could raise output as much as 15-fold in computer science and statistics, but only three- or fourfold in chemistry and biology because laboratory bottlenecks remain. Claude Science offers an early sign that specialised tools can shift usage: ideation reached 8% of its sessions, roughly twice Claude's share, and over 70% of sessions showed high AI autonomy, versus 23.7% on Claude.

Google is rolling out Gemini Omni 1.2 Flash, making it generally available across the Gemini API and Gemini for business. The new version costs $0.07 per second for 720p video, 30% less than its predecessor, and can generate and edit clips of up to 15 seconds, compared with 10 seconds previously. Users can specify durations in one-second increments from three to 15 seconds.

The commercial pitch is as much about workflow as price. Customers can draft at lower resolutions, fix individual details through follow-up instructions and upscale only their best results to 1080p or 4K. That avoids paying to regenerate an entire clip because one product, background or lighting detail is wrong. The model also accepts combinations of character, product and style images, short motion clips and audio cues, with Google saying faces, clothing, packaging typography and brand colours remain consistent across camera angles.

The most striking addition is conversational editing across a full 15-second clip. A user can swap a product or background, change the weather or switch the aspect ratio while preserving the rest of the scene. In one demonstration, successive prompts moved the same two characters through five backgrounds while keeping their performances, armour reflections and camera framing intact. Google also claims more natural dialogue, clearer lip-sync and the ability to assign different voices to the right characters.

Not everything is solved: Google says development continues on fast sports action, rapid script changes and complex physics. The genuinely new elements are the longer native clips, multi-turn editing and broader availability; improvements to consistency and camera control are incremental. Outputs carry an imperceptible SynthID watermark. The model is also reaching Google's own products, including Flow, Google Vids, the Gemini app for AI Ultra subscribers and Google Ads, with 15-second clips in Vids available at no additional cost to users with a Google account or Workspace subscription.

The models sit in Nvidia's Cosmos 3 family and are intended to help developers test machines in virtual environments before putting them on roads or factory floors. Two focus on autonomous driving: one produces driving footage across as many as 11 camera views, with different weather, lighting and times of day, while the other generates successive chunks of camera video and LiDAR, the laser-based sensing used to build a three-dimensional picture of a vehicle's surroundings. The other two turn a physics simulator's controls into realistic video, or predict what a two-armed robot and its cameras would see after a movement.

The practical appeal is the ability to create more training and testing situations without filming or physically staging every one. Researchers can change road layouts and traffic signals, then observe how a driving system responds. Robot developers can start with an image, move a robot or camera, and generate the next view. For the two-armed system, Nvidia supports YAM and AgiBot robots, controlled either by software or by a person using joysticks.

The most useful distinction is where the physics comes from. Sim-Transfer leaves motion, contact and physical state to a conventional simulator, then supplies the colour images; Sim-Bimanual predicts the next observation directly from the scene and the robot's actions. That makes the releases tools for building testing environments, rather than evidence that a robot or driving system is ready for the real world.

What is new is the combination of open model weights, streaming generation and recipes for adapting the systems to a company's own cameras, sensors or robots. Nvidia also warns that errors can accumulate: a slightly mispredicted object becomes part of the context for later frames. Its final training stage therefore exposes the model to its own generated history. The company says researchers still need to establish where the simulations faithfully support policy learning and evaluation, and where they fall short. Running the published setup requires Linux, an Nvidia GPU, Docker and Nvidia's container software; it was tested on one GB200 machine, which Nvidia says is not a minimum-memory specification.

The model, released by Gradium and detailed in a Hugging Face blog post, comes in versions of about 65MB and 75MB. It generates spoken audio entirely on the user's device, with no server and no network connection, so the text being read aloud never leaves the machine. It speaks English, French, German, Spanish and Portuguese, offers more than 100 built-in voices, and can clone a voice from a short recording or create one from a written description.

That combination of small size and quality is the central claim. On Hugging Face's Open TTS Leaderboard, which compares more than 9,000 speech models, Phonon ranks first for intelligibility in German and Spanish and in the top three in every language its makers evaluated, alongside models up to 100 times its size. It also tops the leaderboard's CPU streaming benchmark for time to first audio — how quickly speech starts after text is submitted. The smallest version, Phonon 31m, is released under an Apache 2.0 licence, meaning it can be used free including in commercial products; the larger, five-language version is free only for non-commercial use, with commercial pricing based on user numbers rather than usage.

The launch demos show who this is aimed at: developers building products where connectivity or privacy is a constraint. Examples include an Apple Watch giving coffee-brewing instructions with Wi-Fi switched off, an iPhone counting exercise repetitions with no signal, and a talking robot called Lamp from Autonomous. The Hugging Face team built an integration with its widely used Transformers library, so developers could adopt the model from day one.

What is genuinely new is the packaging rather than the concept: on-device speech and open voice models already exist, but Phonon's makers argue the usual trade-off — small models sound poor, good models do not fit — has been eased. The 200,000 downloads suggest developer interest, though they say nothing yet about whether businesses will ship products with it.

The new feature, called Personal Finance, links accounts through Plaid, a service that connects apps to banks. Anthropic announced it on Thursday and said access is read-only: Claude can look at balances and transactions but cannot move, manage or invest money. It starts rolling out in beta this week to users on the Pro, Max and Ultra plans in the US.

The practical pitch is a monthly spending review without a spreadsheet. A user can ask where their money went, and Claude adds up spending across connected accounts and sorts it into categories such as rent and utilities, food and drink, and shopping. Follow-up questions can test whether that is more than usual or whether anything has been missed. Anthropic says the tool also scans transactions across accounts to flag duplicate charges, unexpected fees and steadily increasing bills, and can answer broader questions such as whether someone is spending more than they earn. It can explain terms such as APR, or annual percentage rate, and calculate how much interest a user has paid over the past year.

The launch is a catch-up move rather than a first. OpenAI's ChatGPT introduced similar personal finance tools in May, also letting users connect financial accounts and ask about spending patterns and long-term planning. What is new for Claude is direct access to a customer's own account data, instead of generic money advice, and Anthropic is explicit about the commercial limits: no ads, no sponsored recommendations, and no credit cards, loans or savings accounts of its own to sell.

The sensitive part is the data handover. Anthropic says Claude's access is limited to each connected bank's name, account name and type, the last four digits of the account number, balances, credit limits and up to two years of transaction history. Conversations in the Personal Finance hub are not used to train its models. For Anthropic, the feature extends a push into specialist markets after Claude for Legal and Claude for Financial Services earlier this year; for customers, the trade is convenience against letting a chatbot see a detailed record of their spending.

Box showed off the feature, called Box Content Retrieval, at its developer conference on Thursday. It builds Nvidia's Nemotron embedding system into the core of Box's product, using embedded metadata to answer queries. The practical change is that an agent can receive the most relevant chunks of documents for a user's query or task, rather than working through each document itself.

Box says users should get faster, more accurate and less expensive queries. Kari Briski, Nvidia's vice president of generative AI software, told TechCrunch she expects queries to be answered four times faster, with 20% fewer errors and 90% fewer tokens used. Tokens are the units of text processed by an AI system, so using fewer of them can reduce the computing work behind a query. The system will be available to all Box Enterprise Advance customers, rolling out early next year.

Box says the feature will be most useful when paired with active agents. It gave lawyers as an example: they should be able to find relevant clauses and precedents more easily. Because Nemotron's embeddings map underlying concepts rather than simply matching keywords, a search tool can surface related legal themes even when different wording or language is used. For businesses testing agents on internal documents, that points to a shift from retrieving whole files towards supplying selected passages for a particular task.

The feature implements Nvidia's Nemotron Embed project, which offers open-weight embedding models for automated retrieval systems. These models use large language models to pre-process data so agents can use it more effectively. Box chief executive Aaron Levie said the approach boosts the relevance of its search technology for agents. Although the Nemotron models are freely available, Box's system also relies heavily on its existing metadata and permission controls, which the source says would make it challenging to recreate from scratch. Levie said Box was "probably packing in five different technologies into one simple API."

The deal, announced Thursday, is expected to close in the fourth quarter of 2026. It will support the development of DBX Agent Cloud, a new storage platform aimed at businesses and developers building AI models and agents. Tigris Data was founded by former Uber engineers who helped develop the ride-hailing company’s storage infrastructure. Its technology automatically positions data closer to the users and computing resources that need it, reducing access times.

DBX Agent Cloud will initially offer object storage as a stand-alone product. Businesses will be able to buy storage independently and use it with existing tools and applications built for Amazon S3, a cloud storage service for storing and retrieving large volumes of data. Dropbox plans to expand the platform with capabilities designed to help AI systems work with documents, images, audio and video. Other planned additions include filesystem features that will allow AI agents to manage files and maintain continuity across tasks and sessions.

Dropbox is also bringing more AI capabilities to its core product. Users will be able to ask AI to perform supported tasks using their files, choose between different AI models, and connect their Dropbox content to third-party tools including ChatGPT, Claude and Muse. The company says the updated experience has already reached more than 70,000 teams. The initial rollout is available to eligible Standard and Advanced customers in the U.S., U.K. and Australia, with broader availability planned for 2027.

Dropbox said it currently has no plans to reduce headcount as part of the acquisition, though it cut 9% of its global workforce last month, according to the Wall Street Journal. Tigris Data raised $25 million in a Series A funding round last year. Dropbox has previously acquired several AI-focused companies, including Reclaim.ai, Promoted.ai and Mobius Labs.

The discussions were described by a person familiar with the matter. Palo Alto Networks and Clutch Security did not respond to requests for comment, and The Information first reported the talks. Clutch is three years old and was last valued at $85 million when it raised a $20 million Series A round in January 2025, led by Lightspeed and SignalFire. It is now approaching $10 million in annual recurring revenue, the yearly income a software supplier expects from subscriptions.

At the proposed price, Clutch would be valued at about 50 times that recurring revenue. The striking part is the gap between its current sales and the price a large security supplier is considering: securing the digital identities and system access of AI agents is being treated as a valuable business before those agents are widely deployed. These agents are software that can carry out tasks inside company systems, so their access needs to be managed and limited like an employee’s.

The valuation fits a broader bout of dealmaking rather than standing alone. Upwind, which uses AI to secure cloud infrastructure, recently rejected an offer from DataDog that valued it at roughly 40 times annual recurring revenue. Cyera bought Oasis Security, another company managing AI agents’ access to systems, for $1 billion, while Cisco bought Astrix for roughly $400 million. Buyers are paying for specialists in a market driven by concern that AI makes it easier for attackers to breach systems.

For Palo Alto Networks, Clutch would be its seventh acquisition this year, according to PitchBook. Its other purchases include Chronosphere for $2.95 billion, Console for $500 million in stock and cash, Embrace for $325 million and Koi for $231 million. The genuinely new element is the emergence of AI-agent access control as a distinct acquisition category; the buying spree itself continues an established strategy of purchasing technology rather than developing it in-house.

The announcement came in Airbnb’s third-quarter shareholder letter on Thursday, alongside more detail about how the agent will work. Chief executive Brian Chesky has previously argued that a conventional chatbot is the wrong interface for travel because people enjoy browsing and anticipating a trip. He told analysts that general-purpose agents are suited to routine tasks such as restaurant reservations, but not necessarily to choosing a destination and a place to stay.

Airbnb’s answer is an agent that assembles a travel-specific display around each question. Instead of relying mainly on text, it will draw on hundreds of components including maps, photos, comparisons, calendars and itineraries, combining them in real time. Chesky described the result as offering the right interface at the right time. The company says its nearly 20 years of matching hosts and guests will inform the system, which is intended to preserve the inspiration and comparison involved in planning rather than simply produce an answer.

The more consequential distinction is what Airbnb can do after a booking. The company says its agent will help guests change reservations, resolve payment problems, obtain refunds, rebook and reach 24-hour safety support — tasks it says an outside agent cannot easily perform alone. Over time, Airbnb expects it to cover homes and hotels as well as services and experiences, two areas that became a bigger focus after last year’s redesign. For rivals and travel businesses, that points to competition over the whole customer relationship, not just the search box.

The genuinely new element is the first-half-2027 consumer rollout and the fuller product description; the underlying strategy is not new. The timing also matters for investors: the news arrived with third-quarter revenue up 19% to $4.87 billion, above Wall Street estimates, and a raised 2026 revenue-growth forecast to the high teens from the mid-teens. Chesky called it Airbnb’s fastest revenue growth in more than three years.

The listing comes almost a year after Figure AI chief executive Brett Adcock predicted Agility would be bankrupt within 12 months. Instead, the Oregon State University spin-out, founded in 2015, is using a special purpose acquisition company, or SPAC, to raise new capital from retail investors and private funding. Chief executive Peggy Johnson said the company has avoided the dancing and backflipping demonstrations associated with rivals, focusing instead on repetitive industrial jobs that are hard to fill with human workers.

Agility’s Digit robots are already deployed with nine customers, including Schaeffler, GXO, Toyota Motor Manufacturing Canada and Mercado Libre, typically moving heavy totes. At one GXO logistics facility, they have handled more than 100,000 totes. But the commercial numbers remain stark: Agility generated just $174,000 of revenue in the first six months of 2026, mostly from related parties, while losing $362 million over the same period. That makes the new funding less a victory lap than a race to turn deployments into a sustainable business.

The immediate opportunity rests on Digit 5, a new model designed to work safely around people rather than only inside an isolated work cell. Agility says it has $300 million in planned contracts from 30 customers for the robot, a pipeline Johnson described as one of the driving factors behind going public. Those deals depend on Digit 5 meeting undisclosed targets for throughput, uptime and placement accuracy. Prototypes have not yet been shown publicly; demonstrations are planned for CES 2027, with broad availability expected by the end of that year.

For prospective customers, the pitch is economic as much as technological. Agility says a robot can save $100,000 a year under a robot-as-a-service arrangement, and potentially more if bought outright. But customers must first spend $500,000 on an on-site proof of concept before a full deployment, and only four had signed up by May 2026. Digit can place heavy loads with 98% accuracy, yet still cannot perform complex manipulation tasks. Johnson said Agility’s 65,000 hours of operational data from real customer facilities should help train the robot, but she would not say the new capital would make the company self-sustaining. Investors are therefore being asked to back a business with real robots in the field, but unproven economics and much of its growth still ahead.