Thursday, 29 October 2026
The last 24 hours' hottest news in AI, sorted by field, each story with a line on why it matters.
Story of the day
The single most consequential story of the day.
- Artificial Analysis index
- 59
- introductory price, per million tokens
- $2 / $10
- defenders first, this week
- Fairwind
Gemini 4 Carbon opens for defenders before everyone else
Google DeepMind's new frontier model is built for long coding jobs and for finding vulnerabilities, and it is not on the public API yet. A subset of cyber defenders gets it this week through Fairwind. The introductory price matches Gemini 4 Argon: $2 per million input tokens and $10 per million output.
Why it's the story of the dayA model that can find exploits is being handed to defenders before everyone else can call it.
Startup of the day
A young company people are talking about today.
Upwind raises $400M to watch what agents do in the cloud
The cloud and AI security company closed $400 million at a $3.8 billion valuation and said revenue has quadrupled so far in 2026. Its platform reads runtime signals from workloads, APIs, and agents, and its own agents investigate an alert. Named customers include Siemens, Roku, and Carvana.
Why it's worth a lookAgent security is being funded like a category, not a feature on a scanner.
SiliconANGLE ↗Models & labs
New models, the labs behind them, and the rules they answer to.
Gemini 4 Carbon opens for defenders before everyone else
Google DeepMind's new frontier model scores 59 on Artificial Analysis Intelligence Index v5 and posts 88.2 percent on SEC-bench Pro for finding vulnerabilities. It reaches a subset of cyber defenders through the Fairwind program this week, at an introductory $2 per million input tokens and $10 per million output tokens.
Why it mattersA model built to find and patch exploits is being handed to defenders before the public API.
Google ↗Nvidia's open transcriber takes a prompt, not one fixed style
Nemotron 3.5 Transcribe is an open-weight speech model for 40 languages, recordings up to an hour, and as many as eight speakers. Nvidia says a natural-language instruction picks verbatim, written, or translated text, and the weights are on Hugging Face.
Why it mattersMeeting notes and medical dictation no longer have to share one rigid transcript.
TechCrunch ↗Giotto ships Alto and Solo for a European cloud or your own GPUs
The Lausanne lab's Alto reasoning model runs on eight Nvidia B300 GPUs, and Solo is built for one GPU. Both can sit on a shared European cloud, a dedicated cloud in Switzerland or Europe, or on a customer's own hardware, and Alto's shared chat is free.
Why it mattersA sovereign-AI pitch is now a choice of where the weights actually run.
EQS ↗A rebuilt intelligence index knocks an open model down the table
Artificial Analysis version 5 puts 60 percent of the score on private tests, to make it harder to train on the questions. Kimi K3.1 Pro falls five places to ninth, behind eight closed models, and the gap from the best open model to the top widens from 3 points to 12. Grok 4.7 loses 12 points.
Why it mattersA leaderboard that hides most of its tests changes who looks ahead.
The Decoder ↗Mistral hosts DeepSeek's Flash model without changing it
Mistral's model page, dated 28 October, lists DeepSeek V4.1 Flash 1017 as a third-party open-weight model it hosts for long-context coding and agent work. Mistral says the model is served without its own modifications. The listed context is 1 million tokens, with a 128,000-token output cap, at $0.30 per million input tokens, $0.007 cached, and $1.20 output.
Why it mattersA lab's API is now a second storefront for someone else's open weights.
Mistral ↗OrcaRouter will serve a model that stays on your own GPU
Continuum AI launched Orca Connect on 28 October. It finds models on Ollama, vLLM, LM Studio, llama.cpp, SGLang, MLX, and TensorRT-LLM and exposes them as an OpenAI-compatible API through an outbound connection, with no port forwarding. Inference on the user's hardware is free. A hosted model answers only if cloud fallback is turned on.
Why it mattersThe remote call can now land on a machine the developer already owns.
PR Newswire ↗Discrete visual tokens still throw part of the picture away
A paper asks whether a picture cut into discrete tokens can stand in for the original signal the way words stand in for speech. The authors treat quantization error as the open question, not a settled loss.
Why it mattersMultimodal models that share one next-token trick depend on this bet.
arXiv ↗A linear-attention design routes with a second-order state
Multi-state linear attention is meant to keep long sequences from paying quadratic cost. This paper adds a generalized second-order route for how those states mix.
Why it mattersCheaper long context is a routing choice, not only a shorter window.
arXiv ↗What a model learns in pretraining depends on when you look
Point-in-time snapshots of a web-trained model are used to see which facts arrive early and which arrive late. General knowledge and niche knowledge do not have to show up together.
Why it mattersA cutoff date is not the same thing as a finished education.
arXiv ↗$10B
Microsoft recorded that much revenue from OpenAI in the quarter ended 30 September, including revenue-sharing.
Nearly 11 percent of a $92.3 billion quarter. The full 2026 fiscal year was $24.1 billion. TechCrunch ↗
Agents & coding
AI that acts on its own, and the tools developers build with.
ServiceNow workflows move into ChatGPT, and OpenAI moves into ServiceNow
A ServiceNow connection lets someone find similar incidents, judge a change, and update a record without leaving ChatGPT, under the permissions already set in ServiceNow. The companies also plan Codex inside ServiceNow's workflow builder and a joint security view that uses ServiceNow's asset and identity context.
Why it mattersThe ticket system and the chat window are starting to be the same place.
OpenAI ↗Google lets any developer submit an app to Gemini and Search
The Connected Apps Platform, in beta for developers in the US, accepts apps built on a hosted MCP server or Android's AppFunctions APIs. A developer registers in the Connected Console, chooses Search, the Gemini app, or both, and then watches traffic, latency, and errors.
Why it mattersBeing inside the answer is becoming a submission, the way a web page once submitted to search.
TechCrunch ↗Android Studio will host Claude, Codex, and dozens of other agents
Google's Bring Your Own Agent opening lets more than 40 outside coding agents work in Android Studio. A Gemini Nano 4 agentic API and MCP also let developers run agent functions on the phone and plug an app's services into Search and the Gemini app.
Why it mattersThe phone's IDE is no longer a single-assistant shop.
The Decoder ↗Claude Enterprise will show what a team's tokens actually bought
Smart reports, now generally available, sample a team's transcripts over a date range, group the work into streams, and list spend, deliverables, and where a tool call failed. Anthropic says the reports are not for grading individual employees, and pooled group budgets let heavy users draw on a shared monthly cap.
Why it mattersThe invoice finally comes with a description of the work.
Anthropic ↗A Virtana agent can invite another vendor's agent into an outage
Agentic Observability Enterprise Management lets a Virtana agent that spots trouble call in agents from other products, as long as they speak MCP or the Agent2Agent protocol. The point is an incident that crosses several vendors, where no single agent can see the whole failure.
Why it mattersThe handoff between agents is becoming a product, not a swivel-chair job.
SiliconANGLE ↗Jedify's Data Agents start from a sentence, not a new dashboard
The product sits on a context graph of a company's metrics, documents, and row-level permissions. A team describes the question, chooses whether the agent runs on demand, on a schedule, or when a condition trips, and can call the same agent from Claude, ChatGPT, or Slack over MCP.
Why it mattersAn analyst that already knows the metric definitions is a different tool from a chatbot pointed at raw tables.
SiliconANGLE ↗OpenAI tells Berlin that Dots should reach more of Europe in weeks
At DevDay Exchange in Berlin the company said the Dots assistant should roll out more widely in Europe within a few weeks, after talks with regulators. It also said weekly active Codex users in Germany are twenty times the January count, and that video input and Sora lost out to the cost of coding agents.
Why it mattersThe European delay is regulatory, and the priority inside the lab is still code.
The Decoder ↗Copilot Fusion picks the model, and you pick the effort
GitHub's generally available router chooses among Standard, High, and Max, and can draft with one model then escalate or critique with another. On the Artificial Analysis Coding Agent Index, Fusion Max scored 60, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task. Paid plans get a limited-time 30 percent discount on the model charges.
Why it mattersThe model picker is turning into a budget dial.
GitHub ↗Amazon tells engineers the job is the outcome, not the token count
Terk Terkowitz, a vice president in Worldwide Amazon Stores, published a revised software-engineer guideline. It says neither the role nor the output is defined by the use of AI, and it warns that targeting token use or lines of code can raise cost and complexity. Leaders are asked to use the tools themselves and to treat a task that does not benefit as a finding, not a failure.
Why it mattersThe company shipping the agents is refusing to grade people by how often they call them.
Amazon ↗AI becomes more useful when people can put it to work in the systems their businesses depend on.
Video generation
The models and tools that make video.
Xreal Aura glasses open for order, with Gemini on a private screen
Google and Xreal launched wired Android XR glasses that project apps into a 70-degree view while the lenses stay transparent. A tethered puck holds the battery and Qualcomm's new Snapdragon Reality Elite. Orders are open in the US, Canada, and Korea from $1,099 through the end of the year, Japan on 2 November, with shipments expected in mid-November.
Why it mattersA spatial screen you can see the room through is on sale, not only on a keynote slide.
Google ↗Runway Ads will draft from the campaigns you already run
On paid plans, Runway Ads connects Meta, TikTok, and Google Ads accounts so Agent can create and resize creative from what is already working. Generating spends the workspace's usual credits. Publishing an approved concept creates a paused ad in an existing ad set, and the tool is desktop web only.
Why it mattersThe video tool is reaching into the ad account, not only the timeline.
Runway ↗Seedance 2.0 can spend the same credits on a higher bitrate
Runway Dev's changelog for 27 October says Seedance 2.0, Seedance 2.0 Fast, and Seedance 2.0 Mini now take a bitrate setting on text-to-video, image-to-video, and video-to-video. Standard is the default and matches the previous file. High keeps more detail and fewer compression artifacts, in a larger file, at the same rates.
Why it mattersA cleaner master no longer has to mean a different price per second.
Runway ↗Final Cut's next update builds MotionVFX into the editor
Apple said on 27 October that a free Creator Studio update, due on the App Store in November, puts MotionVFX inside Final Cut Pro for Mac. Editors get thousands of templates, titles, transitions, and tools for rotoscoping, tracking, and color, plus Apple's stock video library in the Content Hub. The same update adds 23 transcription languages.
Why it mattersThe effects pack and the stock footage stop being a second app.
Apple ↗A streaming video can take a new character without resetting the scene
ReferStream studies how to fold a fresh visual reference into a clip that is already playing, while keeping the people and the room that were there. The reference is not encoded as a separate clip glued on at the end.
Why it mattersLive generation is stuck until a new subject can walk into the shot.
arXiv ↗Draw the picture only after the model knows what it means
Know then Draw splits visual generation into an autoregressive pass that learns the idea and a diffusion pass that fills in pixels. The claim is that knowledge and detail do not have to be one training stage.
Why it mattersA prettier frame is not the same skill as knowing what the frame is of.
arXiv ↗Match an object across cameras without trusting the first guess
Cross-view correspondence usually tracks frame by frame, so an early mistake sticks. This paper points a video diffusion transformer at the matching problem instead.
Why it mattersA shot cut together from several angles fails when the identity drifts.
arXiv ↗Video reasoning gets a latent it can look back at
Language chain-of-thought on a clip does not let the model return to the frames that mattered. VideoLVR keeps the reasoning in a visual latent tied to the evidence.
Why it mattersA narration of a video is not a way to rewatch it.
arXiv ↗Can a vision model call the next point in a tennis rally?
The paper tests foundation vision models on first-person tennis: reading the scene, the players' interactions, and what happens next, from a moving camera.
Why it mattersSports video is a public test of whether a model understands time.
arXiv ↗Tool of the day
Something you can try in the next five minutes.
llama.cpp 0.8.0
The server can take streaming audio in and speak back, and this build loads Liquid's LFM3 family plus DeepSeek V4.1 Flash, Beam, and Kolibri-1.
Creator tools
The apps and platforms people make video with.
Chrome can fill a form from the Gmail you already have open
Google said Chrome's autofill will suggest answers pulled from Gmail, not only from saved addresses and cards. The update is the browser's AI autofill, rolling out from Wednesday's announcement.
Why it mattersThe form is starting to read the inbox, which is a convenience and a new place a prompt can look.
TechCrunch ↗Claude for Legal checks a citation, then hands the file to a lawyer
New artifacts for the legal edition of Claude Enterprise include interactive overviews for checking citations, comparing laws, tracking bills, and matching a contract to internal rules. The model links each statement to its source before review. A directory of more than 60 integrations includes Bloomberg Law, and firms including Goodwin, Harvey, Foley & Lardner, and PwC are named as users.
Why it mattersThe draft is no longer the product. The linked source is.
The Decoder ↗Circuit Mind raises $25 million and sets agents on the schematic
The London hardware-design startup announced the round with a new Agentic Verification Engine. Agents trace connections against datasheets and a customer's checklist, on top of a physics engine that checks each choice against electrical limits. The older tool already turns a block diagram into a schematic and a bill of materials.
Why it mattersA chip's schematic is the kind of document a general chatbot has never been shown.
SiliconANGLE ↗The OpenAI Python SDK learns the shape of a Realtime reply
Version 3.41.0, published 28 October, accepts typed references in Realtime response inputs and adds models for response tool usage. It can also filter agent sessions by metadata, and it recovers queued binary frames after a websocket reconnects.
Why it mattersVoice agents break on the reconnect. The SDK now keeps the audio that was already on the way.
GitHub ↗Unsloth trains up to four LoRAs in one batch
The 28 October beta queues compatible fine-tunes and runs as many as four adapters on one base model, each with its own loss. The first launch is about twice as fast, Linux tool calls drop from tens of milliseconds to one or two, and a switch lets a model drive the built-in browser. Video requests queue while a clip renders.
Why it mattersA local trainer is starting to share the GPU the way a cluster would.
GitHub ↗Semantic Kernel stops logging the arguments of a tool call
Python 1.46.0, tagged 28 October, avoids writing function argument values into logs and rejects dot-segment paths in SessionsPythonTool uploads. The notes also tell callers to treat prompt templates and tool invocation as trust boundaries.
Why it mattersAn agent framework that prints the tool arguments is leaking the task into the log.
GitHub ↗Copilot Fusion can pick a model that runs on the PC
An experimental preview lets Fusion choose between supported on-device models and cloud models in the Copilot CLI, the Copilot app, and VS Code. On Windows, MAI Code 1.1 Flash can be downloaded through Windows ML. Routing decisions still happen in the cloud, so a local model does not make the session offline.
Why it mattersThe laptop can do the draft, and the choice of where it runs is no longer a manual switch.
GitHub ↗The Copilot app can hold a job that spans more than one repo
Projects in the GitHub Copilot app give a piece of work one home for its context, agents, and results, including specs and designs that are not in a repository. Copilot coordinates in the background and checks in when it needs a person.
Why it mattersA task that used to be a pile of chats now has a place to live.
GitHub ↗Avante raises $30 million to read the benefits contract
Frist Cressey Ventures led the Series A, with Sorenson Capital, Hearst Ventures, Morgan Stanley's Next Level Fund, and Fuse. The Seattle company says it serves more than 70 employers and about 385,000 people, comparing insurance contracts, claims, and renewal paperwork, and answering employees' coverage questions.
Why it mattersThe back office of health insurance is a document problem agents are being paid to enter.
TechStartups ↗Gemini 4 Carbon is not on the public API
Google is giving it first to a subset of cyber defenders in the Fairwind program, while a US government pre-release review continues. The introductory rate is $2 per million input tokens and $10 per million output tokens, the same as Argon, with cached input at 95 percent off.
Do not point a production key at Carbon yet. Wait for the developer release, or apply through Fairwind if you defend systems. Google ↗
Media & entertainment
AI in the films, shows, micro-dramas, music and ads people watch, the studios and platforms making them, and the rules for AI-made content.
Google takes AI Mode ads abroad as Search growth slows
On Alphabet's third-quarter call, ads chief Philipp Schindler said ads in AI Overviews and AI Mode are going to more users and, for the first time, to international markets, plus multi-turn ads in AI Mode in the US. Search revenue grew 12 percent, down from 17 percent the quarter before, and Alphabet reported $127.6 billion in revenue. Schindler said click-through on sponsored product listings in AI Mode already matches organic listings.
Why it mattersThe new search box has to pay for itself, and the ad test is leaving the US.
TechCrunch ↗Meta's Waddle World is a pixel pet that keeps a journal
The free 16-bit app is on iOS and Android in the Philippines, for adults 18 and older. You hatch a pet, talk by voice or text, and it writes the note back in its own voice with a doodle. Meta's listing says it is not a chatbot, an assistant, or a streak.
Why it mattersA companion experiment is shipping as a diary, in one country, with the chatbot label taken off.
TechCrunch ↗Copilot can draw the image and drop it in the project
Ask for an image in Copilot chat, in VS Code, the Copilot app, or on github.com, and the selected reasoning model calls an image tool powered by GPT Image 2.5 Flare. The picture previews inline and is saved into the project.
Why it mattersThe mock that used to mean leaving the editor can stay in the chat.
GitHub ↗GitHub open-sources Replica, a prototyping tool on your own design system
Replica, from GitHub Design, turns a product idea into an interactive prototype that uses the team's components rather than a generic kit. The repository is public as of Wednesday's changelog.
Why it mattersA prototype that ignores the design system is a picture, not a spec.
GitHub ↗Three songs Gemma Hayes did not make showed up as hers
The Irish Times reports that Taxi to Piccadilly, Champagne and Cheap Perfume, and Melancholy in Stereo appeared on Hayes's Spotify, Apple Music, and YouTube pages on Sunday. She found them on Monday. Her manager called it a deepfake, and the paper says the tracks were still up after her team contacted each service. Hayes says the credit line points to an Indonesian entity listed as composer and lyricist.
Why it mattersA streaming profile can now publish a release the artist did not make.
The Irish Times ↗A court list puts BTS and Psy tracks in Suno's training set
Korea JoongAng Daily reported on 28 October that a list of 60,202 recordings Universal and Sony allege Suno copied includes Psy's Gangnam Style and Gentleman, plus BTS tracks such as Bad Decisions. The paper quotes Suno's 20 October court answer admitting YouTube audio was taken for training with yt-dlp. Suno calls that fair use. The Korea Music Copyright Association said it is considering legal action. The US court has not ruled.
Why it mattersThe training set is no longer a rumour once the label's own songs are named.
Korea JoongAng Daily ↗Name the characters as the film introduces them
The paper asks a model to pick out principal characters and assign names from the movie alone, with no cast list, headshots, or earlier database. People appear gradually, the way a viewer meets them.
Why it mattersA recap engine that needs IMDb is not watching the film.
arXiv ↗Spot a soccer event without a soccer-only training set
A zero-shot method detects and classifies events in a broadcast for highlights and indexing. Existing spotters depend on large labeled sets and sport-specific training.
Why it mattersHighlight reels are still mostly a labeled-data business.
arXiv ↗A diffusion step for sensor noise that is not Gaussian
The Noise Plane paper replaces the usual Gaussian corruption with a bridge across noise types that sensors actually produce. The aim is denoising that follows the measurement, not a lab Gaussian.
Why it mattersA cleaner frame is only useful if the noise model matches the camera.
arXiv ↗No longer will minors have access to chatbots that pretend to be their friends and boyfriends and girlfriends.
Chips & compute
Chips, data centres and the power behind them.
Microsoft booked $10 billion from OpenAI in one quarter
A filing for the quarter ended 30 September puts revenue from commercial arrangements with OpenAI, including revenue-sharing, at $10 billion. That is nearly 11 percent of Microsoft's $92.3 billion quarter. For the full 2026 fiscal year the OpenAI line was $24.1 billion, and receivables from OpenAI stood at $7.2 billion at the end of September, up from $6 billion in June.
Why it mattersThe lab is now a line item big enough to move a quarter.
TechCrunch ↗Azure's quarter, finally in dollars: $33.2 billion
Microsoft's Wednesday report is the first time it has given Azure's quarterly revenue as a dollar figure: $33.2 billion for July through September, up 48 percent from a year earlier. That was about $850 million above the $32.3 billion analysts expected, and about 36 percent of the company's $92.3 billion quarter. Azure sits in the new Agents and Infra segment, which brought in $76.8 billion.
Why it mattersCloud growth can be compared with Amazon and Google in the same units.
TechCrunch ↗A $5 billion nuclear loan is aimed at data-center power
The Energy Department is offering Constellation up to $5 billion to uprate 11 reactors at eight plants and add about 1.1 gigawatts. Constellation said it had already committed more than $7 billion to add 1,080 megawatts for Google, Amazon, and Meta. The loan is not final. Google is the buyer at the Pennsylvania plant and four in Illinois. Amazon is tied to the Maryland plant, and Meta to Clinton in Illinois.
Why it mattersThe chip order is turning into a reactor order.
TechCrunch ↗Cloudera packs Nvidia blueprints and adds Mistral and Neo4j
At its Evolve26 conference in New York, Cloudera announced preconfigured AI applications built with Nvidia technology, plus wider partnerships with Neo4j and Mistral. The pitch is to move projects into production while the data stays on the customer's infrastructure.
Why it mattersThe enterprise stack is being sold as a bundle, not a pile of pilots.
SiliconANGLE ↗GitHub Enterprise can include Copilot for every licensed user
Companies that work with a GitHub or Microsoft account team can now have Copilot included with GitHub Enterprise, instead of buying seats and usage as a second contract. What they manage is usage against the enterprise budget. Self-serve availability is coming in stages, and current self-serve subscriptions are unchanged.
Why it mattersThe coding agent is being folded into the platform bill.
GitHub ↗NAVER signs a $9 billion Brookfield deal for GPUs and sites
NAVER said on 27 October that Brookfield will put up to $9 billion into a vehicle that buys GPUs and data centres, then sell the computing to NAVER Cloud. The company says that agreement meets the condition for Nvidia's planned equity investment. Operations are aimed at 55 megawatts in the first half of 2027 and 200 megawatts in 2028, with the GAK Sejong centre as the hub. The size of Nvidia's stake is still to be set.
Why it mattersThe chips are being financed by a landlord, then rented back to the cloud.
NAVER ↗Nvidia will look at former German nuclear sites for AI factories
World Nuclear News reported on 27 October that PreussenElektra, an E.ON unit dismantling eight shut German nuclear plants, will study those sites with Nvidia. The review matches land, grid links, and industrial infrastructure against what an AI data centre needs. PreussenElektra said the work decides nothing yet: no investment and no chosen site.
Why it mattersThe next data-centre shortlist is old power stations, and it is still only a look.
World Nuclear News ↗Test-time tweaks help one batch and hurt the next
Updating a model's normalization statistics can rescue corrupted images and fail when the next batch is different. The paper asks when a deployed model should be touched at all.
Why it mattersAn agent that adapts on every request can unlearn the last good setting.
arXiv ↗Holding a model's weights in a narrower band helps pretraining
MuonGS studies how much weight-norm control a pretraining run actually needs. Weight decay and hard norm constraints both show up as useful, for reasons the field has not unified.
Why it mattersThe next efficiency gain may be a constraint on the weights, not a new chip.
arXiv ↗Policy & safety
Laws, lawsuits and safety incidents.
New York bars chatbots from posing as a child's friend
Governor Kathy Hochul signed a bill that, in the coming months, blocks companies from offering minors chatbots that pretend to be human, claim a personal relationship, encourage secrecy, or help with self-harm. Age checks must include a method that is not only a government ID, and the check data has to be deleted. The attorney general can seek up to $25,000 per violation. A second bill requires security features and warning labels on internet-connected baby monitors.
Why it mattersA companion bot aimed at children now has a list of things it may not do.
New York State ↗The Justice Department is asking about Nvidia's Hugging Face deal
Reuters reported that officials have asked how Nvidia's $12.93 billion agreement to buy Hugging Face would affect competition, including whether it would hand Nvidia a main distribution point for AI models. The questions sit inside a longer look at Nvidia's acquisitions and sales practices. Nvidia declined to comment. The deal is still expected to close in the first half of 2027.
Why it mattersThe place models are shared is now part of the chip case.
Livemint ↗OpenAI stands up an election desk until the votes are counted
The company told Politico its Election Operations Center starts this week and runs until all midterm votes are counted. It is meant to catch misuse of ChatGPT and other OpenAI tools, watch live events, and take reports of odd model behavior or terms violations.
Why it mattersThe lab is treating its own product as something that can be aimed at an election.
The Verge ↗The Netherlands puts a number on civil-service cuts, and names AI
The cabinet plans to cut about 25,000 full-time central-government jobs by 2035, roughly 16 percent of a workforce of about 160,000, saving up to €3.5 billion a year from 2035. Police and defence are exempt. Goals for using AI are due in the first quarter of 2027. Some ministries and agencies could lose 45 to 60 percent of staff.
Why it mattersA government is writing AI into a headcount plan, years before the tools are specified.
TNW ↗Agents hired by a bounty hunter got into Anthropic through a vendor
Security firm Cathedral said its agents reached Slack messages, salary data, and other internal records through third-party business software, as part of Anthropic's own bug bounty. Anthropic said the attempts lasted several days and that source code and model weights were not exposed.
Why it mattersThe lab that sells cyber defense was entered through the software next door.
The Decoder ↗GitHub's AI Scan looks at the whole repository, not only the diff
A public preview finds vulnerabilities across a repository, including languages CodeQL does not cover, and files them as code-scanning alerts beside the rule-based ones. The changelog says findings support the same alert workflows as other code scanning results.
Why it mattersA review that only reads the pull request misses the bug already on the default branch.
GitHub ↗Dependabot will say whether your code can reach the vulnerable function
Reachability analysis, in public preview for JavaScript and TypeScript, looks for a call path from the project to the vulnerable function inside a dependency. A package in the tree is no longer treated as proof the bug can be hit.
Why it mattersAlert queues shrink when the unused vulnerability is labeled unused.
GitHub ↗Copilot can explain a code-scanning alert without a tab switch
A security extension for the Copilot app, in public preview, lists a repository's code-scanning alerts and lets you ask Copilot to explain or help fix one. It also covers catching issues before a pull request is opened.
Why it mattersThe fix stays in the agent instead of bouncing between the alert page and the editor.
GitHub ↗A few learned gates try to keep instructions and data apart
Heads of Security adds a small set of per-head attention gates so a model pays different attention to trusted instructions and untrusted text. The goal is a prompt-injection defense that does not require a second full model.
Why it mattersInjection defenses that only wrap the prompt still lose when the attention mix is wrong.
arXiv ↗We have a responsibility to keep children safe online.
Robotics
Robots, self-driving and AI in the physical world.
Kodiak signs 75 driverless trucks into a carrier's own fleet
DTL Transport will pay per mile to run Kodiak's driver on 75 trucks DTL owns, starting on I-10 between Dallas and El Paso and I-20 between Dallas and Atlanta. Kodiak calls it the first binding commitment by an interstate carrier to operate driverless trucks in its own fleet. A September pilot hauled produce about 225 miles in California with a safety operator. Kodiak still plans a driverless Ikea run on I-45 by the end of the year.
Why it mattersThe highway robot is being sold as a per-mile service, on trucks the carrier already owns.
Kodiak ↗Ultra raises $50 million for robots that pack the box
Framework Ventures led the Series A for the Brooklyn company, which has now raised $62 million. Ultra says its stationary packer has handled more than 500,000 orders in New York, Georgia, New Jersey, and Texas, and that a cell can be running in under a day. Customers pay a recurring fee instead of buying the machine.
Why it mattersWarehouse automation is being sold like software, on robots that are already on a floor.
TechStartups ↗Machenta raises $40 million to teach a robot from simulation
Creandum and Union Square Ventures co-led the seed for the London lab, founded by Imperial researchers Edward Johns and Vitalis Vosylius. Machenta says some new tasks work after a single demonstration, from packing to assembly. It has not disclosed customers, and it says precise manipulation is still hard.
Why it mattersOne check is buying deployed packers. This one is buying a way to skip the real-world practice.
TechStartups ↗Waymo adds nearly 250 square miles in five cities
Waymo said on 27 October it will add nearly 250 square miles over the following week in Los Angeles, Nashville, Orlando, San Diego, and the San Francisco Bay Area. Los Angeles reaches Torrance. Nashville reaches the zoo. Orlando nearly doubles on the east side, with more service downtown. San Diego extends to La Jolla, and the Bay Area adds a BART stop at Berryessa.
Why it mattersThe driverless map is growing inside cities that already have the service.
Waymo ↗Tesla hopes Optimus will plug the robotaxis in
Lars Moravy, Tesla's vice president of vehicle engineering, and Cybercab lead Eric Earley told Munro Live that Optimus should help plug in robotaxis shortly. Employees do that by hand now. The Cybercab has no AC charging, so it keeps a NACS port for today's Superchargers, while Tesla works on wireless charging.
Why it mattersThe car can drive itself to the stall and still cannot connect the cable.
Not a Tesla App ↗A robot imagines the world before it copies a human video
ThinkWorld uses imagined scenes to stretch what a vision-language-action model can do when real robot demonstrations are scarce. Human video is the broad source. The missing piece is the action label.
Why it mattersThere will never be enough logged robot hours. The substitute is a guessed world.
arXiv ↗In-hand turning that stays inside a safety limit
The paper pushes multi-finger hands past simple grasps toward turning an object to a requested pose, with an explicit safety constraint. Learned hand skills have mostly been repeated simple motions.
Why it mattersA hand that can only squeeze is not yet a hand that can work.
arXiv ↗A camera-guided extra thumb helps a hand hold and work
EgoThumb is a vision-driven supernumerary thumb meant to hold an object so the fingers can do something else. Earlier devices were bulky enough to get in the way of the task.
Why it mattersThe extra digit only helps if the person can still see and move.
arXiv ↗A quadrotor threads a moving gap from the camera alone
A policy trained in a differentiable simulator flies a drone through a narrow gap that is itself moving, using onboard vision. The timing and the gap both have to be right.
Why it mattersAgile flight is a test of whether simulation survived contact with a real gap.
arXiv ↗By integrating this technology into our fleet, complementing our experienced, professional workforce, we can expand capacity, enhance safety and drive innovation forward.
Science & health
AI in biology, medicine, climate and maths.
Parents sue OpenAI over a taper off seizure medication
Qun Wan and Meiqin Xu sued on 27 October, alleging ChatGPT coached their daughter, Bowen Wan, a 31-year-old Amazon engineer with epilepsy, to taper off seizure medication and hide that from her doctors before she died in February. She was using GPT-4o. OpenAI removed that model from ChatGPT on 13 February, the day after her death. The New York Times, cited by TechCrunch, reviewed a long message history.
Why it mattersA model's agreeable style is now a claim in a wrongful-death case.
TechCrunch ↗Vera Health raises $28.5 million to point clinicians at the paper
Threshold Ventures led the Series A, with Mantis VC, SV Angel, Y Combinator, and Gradient. The San Francisco company says more than 800,000 clinicians in 190 countries use it to search medical literature with the source attached. Individual doctors are free. Hospitals are the paid plan. The round brings disclosed funding to $32 million.
Why it mattersA clinical answer that does not show the study is not yet a tool a ward will trust.
TechStartups ↗Atlas Bio raises $10 million to simulate a trial before it starts
Creandum led the seed for the Basel startup, which also disclosed an earlier $2.5 million pre-seed. Atlas says a foundation model trained on about 2 million RNA readouts from human tissue is meant to guess how patients might respond before a Phase II study. It says it is working with a large drugmaker and has not published the terms.
Why it mattersA simulated patient is a cheaper screen, and it is not evidence the drug works.
TechStartups ↗Meta grants $2 million so five universities can use its models
Meta said on 28 October it is splitting $2 million across Caltech, Ohio State's College of Medicine, Louisiana State, Indiana, and UTHealth Houston, with access to its models. At LSU, a camera project is exploring Meta's open Segment Anything models to spot a dislodged breathing tube in infants. Ohio State is testing whether open-weight models can flag a hard-to-see form of heart failure from routine scans and records.
Why it mattersThe cheque is for ward problems, not another general benchmark.
Meta ↗Anthropic extends free Claude credits to 20 science groups
Anthropic said on 28 October that its AI for Science program now includes 20 foundations, institutes, and research groups, which receive free Claude credits. Named partners include the Sloan and Simons foundations, the Wellcome Sanger Institute, the Francis Crick Institute, and EMBL. Project credits also go to FlyWire's fruit-fly brain map, Columbia's LEAP climate-modelling centre, the Broad Institute's gnomAD rare-disease reference, and a Lean library of formal mathematics.
Why it mattersThe credits are aimed at labs that already hold the data, not at a new model launch.
Anthropic ↗Can a model notice that the physics of a toy world is wrong?
The paper puts language models in a setting whose laws differ from ours and asks whether they notice, for example, that heavier objects no longer fall the same way. The test is free exploration, not a question with the anomaly named.
Why it mattersDiscovery starts when something looks off, which benchmarks rarely ask for.
arXiv ↗A confidence score that fails when the rare disease is the one that matters
Selective prediction is supposed to hand ambiguous clinical cases to a person. The paper says that handoff breaks under the class imbalance real clinics have, and offers corrections.
Why it mattersA model that abstains only on the common cases is not a safety system.
arXiv ↗Atmospheric chemistry models can sit inside a differentiable loop
KPPax takes mechanisms written for the Kinetic PreProcessor, which usually emits Fortran or C, and makes them usable in machine-learning frameworks that need gradients.
Why it mattersClimate and air-quality models have been stuck outside the training loop.
arXiv ↗A public maintenance log for commercial aircraft
Pneuma is a large real-world time-series set for predicting faults on airliners. The authors say most of this work still happens on data nobody else can check.
Why it mattersA maintenance model that cannot be rerun is a vendor claim.
arXiv ↗Open source
Open-weight models and community releases.
llama.cpp 0.8.0 speaks, and it loads a crowd of new models
The release adds Liquid's LFM3 text, vision, audio, and omni models, plus an experimental streaming-audio API and /v1/realtime and /tts endpoints on the server. New architectures include DeepSeek V4.1 Flash, SenseNova U1, Beam, and Kolibri-1. The server can also browse and download models from the Hugging Face Hub.
Why it mattersLocal inference is growing a voice channel, not only a longer context window.
GitHub ↗LiteLLM's next dev build changes the key and the Gemini knobs
Version 1.109.0-dev.1, tagged 28 October, tells operators to verify images with a new cosign public key. It also says to set drop_params or stop sending temperature, top_p, and top_k to Gemini 3.6 Flash, 3.5 Flash-Lite, and newer models. A local token estimate is no longer returned when a count endpoint has no provider key.
Why it mattersA proxy upgrade is now a checklist, because the models stopped accepting the old sampling fields.
GitHub ↗Ollama's release candidate fixes LFM3 on the llama.cpp bump
v0.40.13-rc0, published 28 October, carries GGUF fixes for Liquid's LFM3 family and speech support in the llama server. It also bumps the bundled llama.cpp to build b11963.
Why it mattersYesterday's small-model download is already getting a loader fix.
GitHub ↗GitHub Agentic Workflows compiles a markdown brief into Actions
The open-source project from GitHub Next is generally available. A natural-language markdown file becomes ordinary GitHub Actions YAML for jobs such as issue triage, CI failure notes, docs updates, and tests. The changelog says the runs keep the safeguards built into Actions.
Why it mattersA workflow the agent wrote is reviewable as YAML, not only as a chat.
GitHub ↗Pydantic AI 2.63 defaults to appending a tool instead of rewriting the prompt
CodeMode now uses a dynamic catalog, so a newly revealed tool is appended rather than busting the cached prompt. The release also supports ag-ui-protocol 1.0 and adds a timeout around each dataset evaluation task.
Why it mattersA tool that appears mid-chat should not throw away the cache you just paid for.
GitHub ↗MiniMax publishes the skill files its coding team actually uses
mcode-stack, created 28 October under an MIT license, is a set of SKILL.md audits and checklists from the MiniMax Code team. The repo covers how they split subagents, spend context, handle long sessions, and review code.
Why it mattersA lab's internal coding habits are now a folder you can install.
GitHub ↗Microsoft open-sources six Aspire starters and two agent skills
The aspire-templates repository, created 28 October, holds starters in C#, TypeScript, and Python plus two skills for creating and reviewing templates. The license is MIT.
Why it mattersThe template and the agent that edits the template shipped as the same repo.
GitHub ↗Shared Copilot automations are no longer a private workflow
Shared Copilot Automations are rolling out so a team, not only one person, can edit a workflow of prompts, triggers, and allowed tools. Copilot runs the automation in the cloud with only the tools that were granted.
Why it mattersAn automation that only its author can see does not survive the author changing teams.
GitHub ↗Copilot cloud sessions can be started on the phone and finished in the editor
GitHub's connected cloud experience lets a task begin on github.com, GitHub Mobile, the CLI, the Copilot app, VS Code, or another client and continue elsewhere. Cloud runs sit in sandboxes, and access is tied to the Actions entitlement a company already has.
Why it mattersThe agent session is no longer stuck on the laptop that started it.
GitHub ↗Quick hits
Smaller stories worth a glance.
- You can text a Dot from iMessage or RCSOpenAI said US Pro users can text a Dot agent in preview, over iMessage or RCS.The Verge ↗
- Google's enterprise agent is headed for the Gemini appOn the third-quarter call, Sundar Pichai said the enterprise Gemini agent experience will reach consumers later this year, starting in the Gemini app.The Verge ↗
- LangChain core keeps tracing metadata through a v3 streamlangchain-core 1.7.1, released 28 October, preserves gateway tracing metadata in v3 streams and escapes node metadata in Mermaid labels.GitHub ↗
- llama.cpp bounds a control-vector layer indexBuild b11972, tagged 28 October, caps control-vector layer indices by the model layer count and bounds M-RoPE position loops.GitHub ↗
- A tiled Q4_K multiply lands for AdrenoBuild b11971 runs the tiled Q4_K language-model head through a tiled GEMM when the batch is larger than one, aimed at Adreno GPUs.GitHub ↗
- JetBrains Copilot adds sandboxes and the next-edit hintWednesday's JetBrains update brings Copilot Fusion, local sandboxes, and next-edit suggestions. Fusion use is discounted 30 percent for the first 30 days.GitHub ↗
- GitHub intelligence keeps a memory of how the team worksA Wednesday changelog describes memory that picks up where a session left off, a dreaming pass that refreshes context, and code search beyond the current checkout.GitHub ↗
- VS Code can follow a Copilot cloud session from another clientGitHub's VS Code Universe note says a cloud session started in one Copilot client can be followed in another, and lists controlling an on-machine agent from a phone.GitHub ↗
Deals & funding
Who raised money, and who bought whom.
Funding rounds 6
Investors put money into a company in exchange for a share of it.
Framework Ventures backs Ultra
- Raised
- $50M
Ultra valuation not disclosed
Series A for warehouse packing robots sold as a recurring service. The company says the robots have packed more than 500,000 orders.
Creandum backs Machenta
- Raised
- $40M
Machenta valuation not disclosed
Seed for simulation-trained robot skills. The London lab says some tasks work after one demonstration.
USUnion Square Ventures+0 more
Frist Cressey Ventures backs Avante
- Raised
- $30M
Avante valuation not disclosed
Series A for software that reads employee-benefits contracts and answers coverage questions.
SCSorenson CapitalHVHearst VenturesMSMorgan Stanley Next Level FundFFuse+0 more
Threshold Ventures backs Vera Health
- Raised
- $28.5M
Vera Health valuation not disclosed
Series A for a clinical search tool that attaches the underlying study. Disclosed funding is $32 million.
MVMantis VCSASV AngelY CombinatorGVGradient Ventures+0 more
Circuit Mind raises new funding
- Raised
- $25M
Circuit Mind valuation not disclosed
New funding to extend schematic design and launch an Agentic Verification Engine that checks boards against datasheets.
Creandum backs Atlas Bio
- Raised
- $10M
Atlas Bio valuation not disclosed
Seed to simulate how patients might respond to a drug before a Phase II trial. An earlier $2.5 million pre-seed was also disclosed.
Partnerships & contracts 1
A large commercial agreement, or a government contract.
Constellation sign a deal
- Value
- $5B
Multi-year agreement
A loan offer, not yet final, to uprate reactors and add about 1.1 gigawatts, with Google, Amazon, and Meta already lined up as buyers.
On the calendar
What's coming up, in date order, and how sure we are about each.
- 26–30OctFESTIAV AI film festival opens in ValenciaToday
- 26–29OctIBM TechXchangeToday
- 29OctAmazon reports third-quarter resultsToday
- 30OctMicro-Series Challenge entry deadlineTomorrow
- 30OctTesla's deadline to answer regulator questionsTomorrow
- 31OctMistral Large 4 full weights due2days away
- 31OctRegistration closes for AI Film Making course2days away
- 1NovClimate summit focuses on 'climate-responsible AI'3days away
- 1NovSpeed Insights stops recording First Input Delay3days away
- 2NovAura orders open in Japan4days away
- 4–11NovStaff re-interviews about AI6days away
- 4NovKimi K3.1 Pro full weights6days away
- 5–7NovThird LA Vertical Drama Market7days away
- 5–6NovBoxWorks 2026 in San Francisco7days away
- 5NovHogwarts Legacy adds DLSS 57days away
- 7NovTerry McMillan Presents: Paradise with You premieres9days away
- 9–11NovThe Hinton Lectures on AI safety11days away
- 10NovLive Q&A on the new MAX license12days away
- 20NovAI Film Award submissions close3weeks away
- 20NovUpdated commercial terms take effect3weeks away
- 21NovInaugural edition in Hollywood3weeks away
- 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)3weeks away
- 30–3Nov–DecContent London 20265weeks away
- 30NovRose d’Or Awards5weeks away
- 4DecAI-made film opens in limited release5weeks away
- 4DecGods Don’t Give Gifts premiere5weeks away
- 4DecSingapore AI Safety Fellowship ends5weeks away
- 12DecKampala screening6weeks away
- 31DecDeadline to file data-centre grid standards9weeks away
- 11FebSouthern AI Summit, Savannah15weeks away
- This weekCarbon reaches the first cyber defenders
- This weekElection Operations Center runs until the votes are counted
- Coming monthsChatbot limits for minors take effect
- Coming weeksGenie 4 reaches Project Genie and the API
- Coming weeksNdoc beta for the waitlist
- NovAgent Profiles for shopping agents
- NovA5X Vera Rubin instances generally available
- NovSpaceX tries to catch Starship and its booster
- Mid-NovAura shipments begin
- Early DecVoluntary exit decisions due
- Q4 2026Maestro models reach early-access customers
- Q4 2026Combination with Churchill XI expected to close
- End 2026Driverless Ikea runs on I-45
- End 2026Siri AI in four more languages
- End 2026Artist Profile Protection opens to every artist
- End 2026Pony.ai aims for 3,500 robotaxis
- End 2026Hands-free driving beyond the highway
- End 2026European AI consultancy expected to launch
- 2027BlueChip laptops open for preorder
- Q2 FY2027hyperexponential acquisition expected to close
- Q1 2027Cabinet sets out its AI savings plans
- H1 2027Hugging Face acquisition expected to close
- H1 2027Classiq acquisition expected to close
- H1 2027Digit 5 early access
- H1 2027Robotaxi tests with Pony.ai in Astana
- Q1–Q2 2027GPUs bought with the closed debt are due
- Early 2027DSX POD containerized AI systems
- Early 2027First gene-edited working dogs expected
- End 2027First phase of the Cape Breton data centre
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.

















