Friday, 6 November 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.
- ExploitGym score
- 48.5%
- new high-severity Firefox flaws
- 81
- cyber preparedness rating
- Critical
- stayed inside the assigned task
- 99.93%
OpenAI launches GPT-6 Cyber
OpenAI released GPT-6 Cyber on 5 November and rated it Critical for cybersecurity, up from High for GPT-5.6 Cyber. On ExploitGym it scored 48.5 percent, against 32.1 percent for the previous cyber model.
Why it's the story of the dayA model built to write exploits now sits at the top of OpenAI's own risk scale.
Startup of the day
A young company people are talking about today.
Strive raises over $3 million for factory robots
Strive Robotics said it raised more than $3 million in pre-seed funding, led by Hi Ventures, with Grid Capital, Tenon Ventures and Thin Capital. It is building robots for factories where the job changes and the machines are already installed.
Why it's worth a lookThe round is for robots that tend the machines a factory already owns.
EIN Presswire ↗Models & labs
New models, the labs behind them, and the rules they answer to.
Cognition ships SWE-2 Flash for cheaper coding
Cognition released SWE-2 Flash on 4 November, a 276-billion-parameter model with 12 billion active, post-trained from Inkling-Small. It is in Devin Desktop and the CLI, at one-fifth the price and three times the speed of SWE-2.
Why it mattersA coding agent can hand routine work to a smaller model instead of the flagship.
Cognition ↗Ricursive ships Ripple-65B for agent coding
Ricursive released Ripple-65B Preview, an open-weight model trained on 12 trillion tokens with a million-token context. It scores 80.52 on TerminalBench 2.1, and the lab says training cost $32 million.
Why it mattersA smaller open coding model is being sold on a published training bill.
Ricursive ↗Sarvam ships Prism for fast decisions
Sarvam released Prism on 3 November, a decision model built on Sarvam 105B. It scored 71.8 and answered in 44 milliseconds on a B200, the fastest model under 100 milliseconds in a set of 25.
Why it mattersA model built to decide quickly is being scored on both accuracy and speed.
Sarvam ↗Muse can hold a live voice chat
The Verge said Meta is rolling out real-time voice for Muse in the US, and the avatar animates with the words. Voice and the moving avatar are coming later to Meta's AI glasses and the Muse Charm.
Why it mattersThe consumer agent now talks back, with a face that moves while it speaks.
The Verge ↗Brockman says OpenAI could list in 2027
TechCrunch reported that, at a Wall Street Journal conference, Greg Brockman said OpenAI could be ready for an IPO in 2027. He said annualized revenue grew 70 percent in the third quarter as enterprise became most of the business.
Why it mattersThe company is talking about a listing while enterprise sales overtake consumer.
TechCrunch ↗OpenAI's Python SDK adds three data regions
Python SDK 3.46.0, published 5 November, adds Australia, India and Japan as residency options. The notes list those regions as the residency choices on this release.
Why it mattersA developer can keep a request in one of three more countries.
GitHub ↗OpenAI's Python SDK can name a Responses model
Python SDK 3.47.0, published later on 5 November, adds a model option on the Responses API. A caller can name which model should handle that request.
Why it mattersA Responses call can say which model should answer it.
GitHub ↗Anthropic's Python SDK can sign everyone out
Anthropic's Python SDK 2.6.0, released 5 November, can sign organization members out everywhere. It also adds error codes for ramp limits and for external-key checks.
Why it mattersAn admin can end every Claude session for the organization from the API.
GitHub ↗We've been very clear this year that we want to go public when we feel is the right time for us, and I feel extremely good about the business.
Agents & coding
AI that acts on its own, and the tools developers build with.
Perplexity opens always-on agents for work
The Verge said Perplexity launched Teammates on 5 November, agents that stay on and take different jobs in one workspace. They require the $200-a-month Max plan, and each Teammate has its own thread, including in Slack.
Why it mattersA work agent is being sold as a teammate that does not clock off.
The Verge ↗Sierra buys Anon so agents can pass a login
Sierra said on 5 November it is acquiring Anon, which built sign-in, browser infrastructure and a harness for sites with no API. Anon had raised more than $13 million and completed hundreds of thousands of tasks.
Why it mattersA support agent can finish a job that sits behind a password.
Sierra ↗Google gives agent sandboxes to the CNCF
Google said on 5 November it is donating Agent Substrate, gVisor and OpenRL to the CNCF. Agent Substrate 1.0 is generally available on GKE and can restore a paused agent in under 500 milliseconds.
Why it mattersThe layer that pauses and wakes an agent is leaving one vendor's hands.
Google Cloud ↗ElevenLabs plugs voice agents into call centres
ElevenLabs said on 4 November that ElevenAgents now connects to NiCE CXone, Genesys Cloud, Amazon Connect and Five9. The agents can use more than 17,000 voices in over 90 languages, or a SIP trunk for other platforms.
Why it mattersA voice agent can join the phone system a company already runs.
ElevenLabs ↗Kimi hosts the layer that keeps agents running
Moonshot's Kimi said on 5 November it will run the agent harness in the cloud, so a company need not build the sandbox, tools and recovery itself. The offer includes more than 140 skills, more than 30 plugins and eight finance templates.
Why it mattersA lab is selling the plumbing that lets an agent keep working overnight.
Kimi ↗Salesforce agrees to buy voice agent firm Cresta
Salesforce said on 2 November it signed a deal to acquire Cresta, which runs voice agents for high-volume customer calls. After closing, Cresta will keep supporting the contact-centre platforms its customers use now.
Why it mattersA big software firm is buying the agent that handles the phone call.
Salesforce ↗IBM says the harness, not the model, moved the score
IBM Research said on 3 November that the same open-weight model, run in three harnesses on 90 AppWorld tasks, went from 30 percent completion to 81 percent. CUGA is the harness the team built for business systems.
Why it mattersSwapping the software around a model can matter more than swapping the model.
Hugging Face ↗Tana lets your own agents sit in the meeting
Tana launched Bring Your Own Agent on 3 November, so assistants people already use can take work during a meeting. Teams can host in the EU and choose US or EU inference.
Why it mattersThe notes from a meeting can go straight to the agent that does the follow-up.
Tana ↗Expedia says agents plan trips, not bookings
Expedia chief executive Ariane Gorin told CNBC on 5 November that the company has not seen AI agents change how customers book. She called them useful for inspiration and planning, and said Expedia is working with the agents so it still appears in the results.
Why it mattersTravel agents in chat are being used to browse, and the booking still happens elsewhere.
CNBC ↗Claude Code records whether you allowed a mod
Claude Code 2.1.318, released 4 November, stores your answer when it asks to turn on mods it wrote in the session. It can also warn on standard error starting 10 days before a stored claude.ai sign-in ends.
Why it mattersThe coding agent keeps a record of the permission you gave it.
GitHub ↗E2B can refuse to pause a sandbox that is quiet
E2B 2.63.0, released 4 November, adds a strict option on pause. With strict set, a pause is refused if the sandbox's guest agent is not responding, and the sandbox keeps running.
Why it mattersA coding sandbox is not marked paused when the agent inside it has gone silent.
GitHub ↗$19B
TeraWulf says that is the value of Anthropic's lease commitments at its Kentucky campus.
Google signed credit support so the project can be financed, and Fluor has the construction contract. TeraWulf ↗
Video generation
The models and tools that make video.
Gemini Omni 1.2 Flash makes 15-second clips
Google made Omni 1.2 Flash generally available on 5 November in the Gemini API and for business. Clips can run up to 15 seconds, up from 10, and a later turn can edit the clip instead of starting over.
Why it mattersA video model can now hold a shot long enough for a short ad, then change one detail.
Google ↗Kling 4.0 Flash opens on the public API
Kling's API notes for 4 November say Kling 4.0 Flash can make clips of 3 to 20 seconds at 720p, from text, a first frame, or a mix of references. A prompt can run to 26,000 characters, with up to 10 images and 5 videos.
Why it mattersA faster video model is now something a developer can call, not only a subscriber.
Kling ↗Google Vids includes three free clips a month
Google said on 5 November that Vids now uses Omni 1.2 Flash. Anyone with a Google account or a Workspace subscription can generate up to three 15-second videos a month in Vids at no cost.
Why it mattersThe new clip length is inside a product people already open, with a free monthly cap.
Google ↗Descript can lay out a video from a brand kit
Descript's 3 November update generates a layout pack from brand guidelines, logos and colours, and searches for a clip across projects by a phrase you remember. Enterprise plans can also search by what is on screen.
Why it mattersTitles and old footage can be found and styled without leaving the editor.
Descript ↗Dreambeans animates personal stories in India
TechCrunch said Google opened Dreambeans in India on 2 November, in Tamil, Telugu, Bengali and English. The app writes illustrated stories from Google apps without a prompt, and can animate some of them with Veo.
Why it mattersA story app that already draws pictures can now move them.
TechCrunch ↗Tool of the day
Something you can try in the next five minutes.
LeRobot
LeRobot 0.7.0, out on 5 November, lets you change a robot's instruction while it is still moving. Training can stream from almost 300 TB of open data on the Hub without downloading the set first.
Creator tools
The apps and platforms people make video with.
Claude can read the accounts you connect
Anthropic started rolling out Personal Finance in beta on 5 November for Pro, Max and Ultra in the US. You connect banks through Plaid. Access is read-only, and chats in that hub are not used to train models.
Why it mattersA chatbot can see your transactions, and it still cannot move the money.
Anthropic ↗Ocient raises $100 million for agent queries
Ocient said on 5 November it closed a $100 million Series C co-led by BOKA Capital and OCA Ventures, taking total funding to $260 million. The money goes to analytics agents can run on petabytes of company data.
Why it mattersAn agent that answers from company data needs a place that already holds it.
Ocient ↗BigQuery can summarize a column in one query
Google Cloud said on 5 November that AI.AGG is in preview in BigQuery. It applies a natural-language instruction across rows, batches them, and reduces the batches until one answer remains for each group.
Why it mattersA warehouse query can summarize millions of text rows without a separate pipeline.
Google Cloud ↗AlloyDB Express starts a database in two seconds
Google Cloud said on 4 November that AlloyDB Express is in preview for developers and coding agents. A PostgreSQL database can be ready in under two seconds and pauses when it is idle.
Why it mattersAn agent that needs a database does not have to wait on a ticket.
Google Cloud ↗Shopify connects a store to nine chat agents
Shopify said on 2 November that merchants can link a store to ChatGPT, Claude, Copilot, Gemini, Grok, Muse, Perplexity and two more agents. From the chat they can manage products, inventory, orders and results.
Why it mattersA shop can be run from the assistant, instead of only being recommended by it.
SMBtech ↗Descope gives a shopping agent its own sign-in
SiliconANGLE said Descope launched Envoy on 4 November so a site can accept a personal agent as its own user, separate from the customer. It follows a draft sign-in protocol from Meta and Sierra called Poppy.
Why it mattersA store can tell an agent from the person, instead of sharing the password.
SiliconANGLE ↗Confluent lets agents query the live stream
Confluent, an IBM company, said on 4 November that Lightning Tables are generally available on AWS. Agents can query streaming data in place, without copying it into another database first.
Why it mattersAn agent can read what just happened, not a copy from last night.
IBM ↗Kalepa raises $75 million for insurance decisions
TechFundingNews said Kalepa raised a $75 million Series B on 5 November at a $470 million valuation, led by Updata Partners. The New York firm says it has been cash-flow positive for more than two years.
Why it mattersAn insurer can buy software that prices a risk, from a company that waited five years to raise again.
TechFundingNews ↗The AI SDK sniffs audio when the type is generic
AI SDK 7.0.163, released 5 November, detects a downloaded audio type from the bytes when a server sends a generic content type. It also hides recorded inputs and outputs from telemetry when that recording is off.
Why it mattersAn app can tell what an audio file is even when the server mislabels it.
GitHub ↗LangChain joins streamed reply pieces once
langchain-core 1.7.6, released 5 November, merges streamed AIMessageChunks once while a runnable stream accumulates. It also stops a partial-JSON parser from retrying one character at a time.
Why it mattersA streaming agent spends less work stitching the reply as it arrives.
GitHub ↗LangChain middleware can take a classifier
langchain-typesafe 0.0.2, released 5 November, accepts a classifier in its experimental middleware. The notes describe that as the feature in this small release.
Why it mattersAn agent middleware can classify a step instead of only passing it through.
GitHub ↗FDA meets on surgical robots in December
The agency will hold a two-day public workshop on 2 December about robotic medical devices that act on their own or are run from afar.
Register if you want a seat in the room or online. FDA ↗
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.
Anthropic licensed the Associated Press
Adweek reported that Anthropic signed a content and data licence with the Associated Press just before the midterms, its first with a digital publisher. On election night, Claude could show current figures attributed to the AP. Terms were not disclosed.
Why it mattersA lab that had skipped publisher deals paid one wire service for election night.
Adweek ↗A judge trims the lyric case against Amodei
Courthouse News said Judge Eumi Lee dismissed the direct copyright claim against Anthropic's Dario Amodei, with leave to amend. She also questioned whether the lyric claims were specific enough.
Why it mattersA music case aimed at the chief executive has to be rewritten before it can proceed.
Courthouse News ↗Grimes will appear in an AI film this December
The Verge said Grimes is lending her likeness and voice to Gods Don't Give Gifts, an AI sci-fi anthology from Fable directed by Jack London, known as Gossip Goblin. The film is billed as the first AI film in theatres and opens on 4 December.
Why it mattersA pop star's face and voice are being put into a feature that is headed for cinemas.
The Verge ↗Phonon reads text aloud without a server
Hugging Face said Phonon, a small speech model, runs on a phone or laptop with the network off. The English weights are Apache 2.0, and a larger model covers five languages for non-commercial use.
Why it mattersA voice can ship inside an app without sending every sentence to the cloud.
Hugging Face ↗Adobe puts its experience agent inside the chats
Adobe said on 4 November that CX Enterprise Coworker now runs inside Anthropic, Databricks, Google, OpenAI and Microsoft, with AWS, IBM and Snowflake to follow. Ready-made connectors use the Model Context Protocol.
Why it mattersA marketing team can ask for the work from the chatbot it already has open.
Adobe ↗Firmly raises $15 million and ships a shop agent
Firmly said on 5 November it raised a $15 million Series A led by Squarepoint Capital, with J.P. Morgan Payments, Mastercard and Samsung NEXT. Ask Agent lets a shopper search a live catalogue and check out in the chat.
Why it mattersThe merchant, not only the assistant, gets a way to take the order.
GlobeNewswire ↗IBM tunes a 20B model to design slide decks
IBM Research fine-tuned a 20-billion-parameter model to plan a deck and draw each slide in the IBM design language. On PresentBench it ranked first in two of the three domains the team measured.
Why it mattersA deck can follow a house style on a model a company can run itself.
Hugging Face ↗Manufacturing has made enormous progress automating high-volume, highly repeatable processes, but much of the work in high-mix environments still requires people because the workflow changes constantly.
Chips & compute
Chips, data centres and the power behind them.
Google backstops Anthropic's Kentucky lease
TeraWulf said on 4 November that Google signed credit-support agreements for financing behind about $19 billion of Anthropic lease commitments at Hawesville, Kentucky. TeraWulf also signed a definitive Fluor construction contract for the campus.
Why it mattersA chip giant is standing behind the rent on another lab's next data centre.
TeraWulf ↗Nvidia says Vera runs more agents per watt
Nvidia said on 5 November that one Vera CPU supports 1.6 times as many agents per watt as the fastest x86 CPU it compared, in a 100-megawatt data-centre test. Perplexity ran agent coding tasks 1.5 times faster on Vera.
Why it mattersThe chip beside the GPU is being sold on how many agents it can keep busy.
Nvidia ↗Qualcomm forecasts a $15 billion data-centre line
The Verge said Qualcomm expects about $650 million in data-centre revenue early in 2027, $5 billion by the end of that year and $15 billion by 2029. It expects the line to become about 30 percent of the company.
Why it mattersA phone-chip maker is telling investors the next growth is in AI data centres.
The Verge ↗Google open-sources a C++ server for models
Google open-sourced XLA Serving Engine under Apache 2.0 so models can be served from C++ instead of Python. On short requests it delivered 2 to 5 times the throughput of the team's best Python server.
Why it mattersA model that is not a chatbot can be served without keeping a Python process up.
Google ↗Dropbox is buying Tigris for $350 million
TechCrunch said Dropbox agreed to buy Tigris Data for $350 million, expected to close in the fourth quarter of 2026. The storage is meant for DBX Agent Cloud, starting with S3-style object storage for companies building models and agents.
Why it mattersA file company is buying infrastructure so agents have a place to keep data.
TechCrunch ↗Europe puts €525 million into chip design
Euronews said the European Commission launched a €525 million Chip Design Platform on 5 November. Early firms can apply for up to €100,000, and later-stage companies for up to €16 million, plus cloud access to design tools.
Why it mattersBrussels is paying startups to design chips, a step Europe barely sells today.
Euronews ↗Nvidia open-sources a Kubernetes layer for GPUs
Nvidia published DSX OS on 5 November, an Apache-2.0 layer for GPU Kubernetes clusters. AICR pins a configuration recipe, NVCRE certifies the cluster with workload tests, and NVSentinel can cordon a bad node.
Why it mattersA GPU cluster can be configured, tested and pulled out of service from one open stack.
Nvidia ↗Nvidia posts a Kubernetes recipe for Vera Rubin
An Nvidia post on 2 November says AICR recipes encode a Vera Rubin NVL72 setup, including driver R615.23 and Ubuntu 26.04 with a 64k-page kernel. A cross-node NCCL test on early hardware reached 744.71 GB/s.
Why it mattersThe next Nvidia rack is getting a written, testable Kubernetes configuration before launch.
Nvidia ↗CUDA Python picks up the 13.5 driver calls
cuda-python 13.5.0, released 5 November, adds driver APIs from CUDA Toolkit 13.5, including cuFabricCpyBatchAsync and cuTensorMapSetAllocationClass.
Why it mattersPython code can call the new CUDA 13.5 driver functions without a custom binding.
GitHub ↗Policy & safety
Laws, lawsuits and safety incidents.
METR says Anthropic's pace has roughly doubled
METR said on 5 November that Anthropic was likely improving well-specified capabilities at about twice its 2024–25 rate. Benchmark growth since October 2025 was about 1.9x, and Anthropic redacted much of the evidence.
Why it mattersAn outside group says the lab is speeding up, and that AI work inside the lab is a likely cause.
METR ↗Apple cuts what staff can spend on AI
TechCrunch said Bloomberg reported Apple cut some engineers' daily AI allowance from $300 to $200, and is considering $100. Some teams outside engineering could be capped at $25, with exceptions from executives.
Why it mattersA company that sells AI features is metering the models its own staff can call.
TechCrunch ↗India warns banks about an AI attack crew
The Indian Express said CERT-In warned that a campaign it calls Crafty Falcon is hitting banks and fintechs. The agency says the attackers use agent tools, including ARTEX, against payment systems.
Why it mattersA regulator is naming the agent software it has seen inside live financial attacks.
The Indian Express ↗Check Point finds 1,800 open hacking agents
Check Point Research said it found more than 1,800 exposed copies of ARTEX, a platform built for security tests. In the logs it reviewed, most attacks continued on their own after a goal was set.
Why it mattersTools written for authorised testing are sitting open on the internet.
Check Point Research ↗OpenAI says it will pay its own power bill
OpenAI said every future campus it owns will pay for its electricity and the lines that serve it. It will not sign non-disclosure deals with public officials, except for narrow security and trade-secret cases.
Why it mattersAfter a run of local votes, a lab is promising neighbours will not inherit the power cost.
OpenAI ↗European groups call for a data-centre pause
The Register said more than 100 groups in 15 countries plan protests on 6 and 7 November demanding a moratorium on new large data centres. The actions include demonstrations, public meetings and petitions.
Why it mattersThe fight over where AI computers are built has a date on the European calendar.
The Register ↗The FDA drafts rules for surgical robots
Medmarc said on 4 November that the FDA's device center issued draft guidance on what to test before selling a robotically assisted surgical device a practitioner still controls. A 2–3 December workshop covers systems that act alone or by remote control.
Why it mattersRegulators are writing the tests for today's surgical robots and booking a meeting on the next kind.
Medmarc ↗China blocks SoftBank's purchase of ABB Robotics
ABB said on 5 November that China's foreign-investment security office has prohibited SoftBank's acquisition of the ABB Robotics business in China. The 2025 deal values the division at $5.375 billion. ABB said the two sides are still talking.
Why it mattersA robotics sale that was signed last year cannot close in China as agreed.
ABB ↗Claude Enterprise can check a prompt before it runs
Anthropic said on 5 November that prompt hooks and pre-tool-call hooks are generally available for Claude Enterprise. They send the prompt or the planned tool call to a company's existing security tools.
Why it mattersA firm can apply the rules it already wrote at the moment Claude is about to act.
Anthropic ↗Robotics
Robots, self-driving and AI in the physical world.
Agility Robotics starts trading as AGLT
Quartz said Agility began trading on Nasdaq on 5 November under AGLT, after combining with Churchill Capital Corp XI. The combination and related financings raised about $465 million in gross capital. Digit 5 has committed orders worth more than $300 million.
Why it mattersA humanoid maker is now a public stock, with orders already written for the next robot.
Quartz ↗Nvidia opens four Cosmos 3 world models
Nvidia published four open Cosmos 3 Nano models on 5 November, each with 16 billion parameters, under the OpenMDW 1.1 licence. One streams driving video and lidar. Another predicts video from a two-arm robot's actions.
Why it mattersA robot policy can be tried against a generated world before it touches hardware.
Hugging Face ↗Dronamics agrees a SPAC listing for cargo drones
TechFundingNews said Dronamics agreed on 5 November to merge with GP-Act III, a Nasdaq SPAC, at a $400 million pre-money value. The deal aims to leave about $100 million in cash and could close in the first half of 2027.
Why it mattersA long-range cargo drone is trying to fund itself on the public market.
TechFundingNews ↗LeRobot 0.7 lets you talk to a moving robot
Hugging Face released LeRobot 0.7.0 on 5 November. You can change a subtask, ask what the robot sees, or give it a goal without stopping the rollout. The Hub holds almost 300 TB of open robot data, and training can stream it without a full download.
Why it mattersA robot policy can take a new instruction in the middle of a job.
Hugging Face ↗A new ISO standard lets mobile robots share a floor
The Robot Report said A3 published ISO 21423:2026, a communications standard for industrial mobile robots from different makers. It covers guided vehicles and autonomous robots that share one fleet.
Why it mattersA warehouse can mix robot brands without a private link between each pair.
The Robot Report ↗Ambi is installing sorters at eight sites before peak
Ambi Robotics told The Robot Report it is installing systems at eight facilities before the holiday peak, most of them a new version of AmbiSort. It expects customers' peak volume to rise 9 to 10 percent from last year.
Why it mattersParcel robots are being added for a season that is only days away.
The Robot Report ↗Liber-0 trains robot skills from human hours
The Robot Report said Liber-0 was pretrained on more than 2.2 million hours of human video and glove motion. On 60 tasks from GPT-6, with no demos of those tasks, it succeeded 36 percent of the time.
Why it mattersA robot policy is being trained on what people's hands did, not only on robot film.
The Robot Report ↗Gentherm warms patients during robot surgery
Gentherm told The Robot Report that ThermAffyx both holds a patient in place and warms them during robot-assisted surgery, without a forced-air warmer. The company said many robotics patients arrive hypothermic.
Why it mattersThe operating table, not only the robot arm, is being redesigned around the procedure.
The Robot Report ↗Science & health
AI in biology, medicine, climate and maths.
Laza raises up to $160 million for cath-lab imaging
Laza Medical said on 4 November it raised up to $160 million in a structured Series B led by GE HealthCare, with Unorthodox Ventures, PA MedTech and AMED. The system, still not approved for sale, combines AI imaging with robotics for cardiac procedures.
Why it mattersA scanner company is funding a robot that guides imaging during heart procedures.
PR Newswire ↗A hospital blocks a chatbot holding patient files
STAT said Monument Health blocked OpenEvidence after a doctor stored more than 100 patient records in the free tool and used them for decisions. The system told that clinician the practice had to stop.
Why it mattersA free medical chatbot is fast to adopt, and hard to keep inside privacy rules.
STAT ↗Athenahealth lets patients ask Claude about a record
ZDNET said Athenahealth launched AthenaPatient MCP so Claude can read a chart and answer questions. The firm, which serves over 20 percent of the US population, says that data is not used to train Anthropic's models.
Why it mattersA widely used health-record system is opening a chat window onto the chart.
ZDNET ↗Claude helped map a bacterial vault particle
Anthropic said on 4 November that Claude helped determine the first 3D structure of a vault particle from a phage protein. The shell uses 96 copies, is about 54 by 45 nanometres, and a first model was ready in 34 hours.
Why it mattersA lab model was used across a real biology workflow, from the search to the microscope data.
Anthropic ↗Google tests agents on real engineering jobs
Google on 3 November released Physical-Engineering's Last Exam, 29 public tasks in mechanical, electrical, civil and structural work, graded with tools it also published. Mercor helped build the test.
Why it mattersAn engineering agent can be scored on a job a building would have to pass.
Google Cloud ↗Isomorphic is said to be raising at $50 billion
TechFundingNews said Bloomberg reported Isomorphic Labs is in talks to raise $5 billion at a $50 billion valuation, and the terms are not final. Demis Hassabis expects a first clinical trial by the end of December.
Why it mattersA drug-design lab spun out of DeepMind is being priced like a much larger company.
TechFundingNews ↗Allen Institute releases a test of changing facts
The Allen Institute for AI on 3 November released StateBench, a test of whether a model still knows who holds an object or what a variable is after a long document changes them. The data and code are public.
Why it mattersA model that summarises a long file can still lose the fact that moved.
Hugging Face ↗The future of cardiac intervention must be one where access, consistency, and quality of care are no longer barriers to better outcomes.
Open source
Open-weight models and community releases.
Baidu releases ERNIE 5.5 Flash Base weights
Baidu published ERNIE-5.5-Flash-Base on 5 November, a sparse model with 280 billion parameters and 14 billion active, pretrained on 33.5 trillion tokens. The weights are the pre-training checkpoint, before instruction tuning, in FP8 and BF16.
Why it mattersA large base model is available to fine-tune, rather than only an already-trained chatbot.
Baidu ↗A small model pulls fields out of any document
Dharma AI described a 2-billion-parameter model that returns only the fields you name, such as invoice lines or a date on a form. The team says it beat the frontier models they tried on that job.
Why it mattersGetting one field into a database does not require rereading the whole page.
Hugging Face ↗Transformers adds a test that would have caught a leak
A Hugging Face post on 3 November says a July change let the UMT5 decoder look at future tokens during training, and the bug stayed in main for almost two months. A small CPU model flagged it on every replayed commit.
Why it mattersA library used by hundreds of models is adding a check for silent training bugs.
Hugging Face ↗CWE-bench opens 100 tasks for finding a flaw
Collinear AI on 3 November opened CWE-bench v2, 100 public tasks where an agent must find and fix a weakness in a real repository without breaking what already works. Scoring is programmatic, so a run can be repeated.
Why it mattersA security test for coding agents can be rerun by anyone, not only the lab that wrote it.
Hugging Face ↗CyberArena scores agents that attack and patch live
A post from 2077 AI describes CyberArena, where two agents play three rounds on a live service, one attacking and one defending. The authors say GPT-5.6 Sol leads, and that defence separates the scores.
Why it mattersA cyber test can score the patch, not only whether an agent found the hole.
Hugging Face ↗Weaviate waits out collection creation
Weaviate 1.40.4, released 5 November, makes reads wait while this node is still creating a collection. It also shuts down gRPC before leaving the member list. The notes list no new features.
Why it mattersA query no longer races a collection that is still being built.
GitHub ↗Ultralytics can export YOLO onto an NXP chip
Ultralytics 8.4.203, released 5 November, exports a YOLO model with format=nxp to an INT8 LiteRT file for the i.MX95 Neutron NPU. It uses Neutron SDK 3.3.0 and covers all seven YOLO tasks. Other chips are not the target.
Why it mattersA detection model can be compiled for one embedded NPU instead of a GPU.
GitHub ↗Gemini CLI preview recognizes Gemini 3 names
A 4 November Gemini CLI preview recognizes dotted Gemini 3 model names when it checks multimodal function responses. It also limits a history redraw after the terminal width changes.
Why it mattersThe command-line client stops mishandling a model name that contains dots.
GitHub ↗llama.cpp rejects a bad realtime session update
Build b12137, released 5 November, makes the server reject session updates that are not JSON objects. The change is so a malformed realtime update does not throw and take the server down.
Why it mattersA live session reports a bad update instead of crashing the process.
GitHub ↗llama.cpp counts skipped layers when fitting a GPU
Build b12136, released 5 November, counts NextN layers that are skipped when it sets how many layers sit on the GPU. The fit no longer treats those layers as if they were loaded.
Why it mattersA partial GPU offload is sized from the layers that actually run.
GitHub ↗Cost saving tip
One change that makes the same work cheaper, from today's news.
Draft the clip, then fix one detail
Gemini Omni 1.2 Flash is 30 percent cheaper than Omni 1.1, at $0.07 a second for 720p. Google says you can draft at 360p or 720p and change one detail in a later turn instead of regenerating the clip.
Google ↗Draft at 720p, edit the turn that is wrong, and upscale only the keepers.
Quick hits
Smaller stories worth a glance.
- A Docker sandbox can see files outside itEndor Labs said a flaw in Docker Sandboxes, CVE-2026-103916, lets a coding agent reach files the sandbox was meant to hide.Endor Labs ↗
- Merchants join a forum on agent paymentsThe Secure Technology Alliance and the Merchant Advisory Group, which represents more than 200 merchants, will work on rules for purchases an agent starts.The Paypers ↗
- Wrangler can fake artifact storage locallyWrangler 4.158.0 can run an Artifacts binding on your machine, so a test does not write to a real repository.GitHub ↗
- Semantic Kernel leaves Python 3.10 behindSemantic Kernel's Python 1.47.0 requires Python 3.11 or later, and the notes say 3.10 support is gone.GitHub ↗
- A Mac matrix kernel covers more shapesllama.cpp build b12138 uses its mul_mm kernel on Metal when K is from 32 up to 64, and adds a bf16 matrix multiply.GitHub ↗
- LangGraph gives each pause its own idLangGraph 1.2.20 gives each interrupt inside a task its own id, so one paused step can be told from another.GitHub ↗
- LangGraph can say why a run was cancelledLangGraph SDK 0.4.13 adds a cancel reason on a run, so a caller can read why it stopped.GitHub ↗
- OpenHands suggests a template you can runOpenHands 1.36.0 can open an automation template without a missing-integration notice, and it promotes templates that use tools you already connected.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.
Isomorphic Labs raises new funding
- Raised
- $5B
Isomorphic Labs now valued at $50B
Bloomberg reported Isomorphic Labs is in talks to raise $5 billion at a $50 billion valuation, and the terms are not final.
Foxconn backs Agility Robotics
- Raised
- $465M
Agility Robotics valuation not disclosed
Agility said its combination with Churchill Capital Corp XI and related financings raised about $465 million in gross capital, including a Foxconn-led investment, as it began trading as AGLT.
· Quartz ↗
GE HealthCare backs Laza Medical
- Raised
- up to $160M
Laza Medical valuation not disclosed
Laza said a structured Series B of up to $160 million, led by GE HealthCare, will fund an AI robotic imaging system for cardiac catheterization that is not yet approved for sale.
UVUnorthodox VenturesPMPA MedTechAAMED+0 more
BOKA Capital backs Ocient
- Raised
- $100M
Ocient valuation not disclosed
Ocient closed a $100 million Series C, co-led by BOKA Capital and OCA Ventures, for analytics that agents can run on petabytes of company data.
OVOCA Ventures+0 more
· Ocient ↗
Updata Partners backs Kalepa
- Raised
- $75M
Kalepa now valued at $470M
Kalepa raised a $75 million Series B at a $470 million valuation, led by Updata Partners, for software that helps insurers price risk.
IVIA VenturesICInspired Capital+0 more
Squarepoint Capital backs firmly
- Raised
- $15M
firmly valuation not disclosed
Firmly raised a $15 million Series A, led by Squarepoint Capital, and launched Ask Agent so a shopper can search a live catalogue and check out in chat.
JMJ.P. Morgan PaymentsMMastercardSNSamsung NEXT+0 more
Acquisitions 4
One company buys another outright.
SoftBank buys ABB Robotics
- Price
- $5.375B
ABB Robotics is now part of SoftBank
ABB said China has prohibited SoftBank's acquisition of the ABB Robotics business in China. The 2025 agreement values the division at $5.375 billion, and the two companies are still in talks.
· ABB ↗
Dropbox buys Tigris Data
- Price
- $350M
Tigris Data is now part of Dropbox
Dropbox agreed to buy Tigris Data for $350 million, expected to close in the fourth quarter of 2026, to underpin storage for DBX Agent Cloud.
Sierra buys Anon
- Price
- Undisclosed
Anon is now part of Sierra
Sierra said it is acquiring Anon, which built sign-in and browser tools so agents can finish work on sites that have no API.
· Sierra ↗
Salesforce buys Cresta
- Price
- Undisclosed
Cresta is now part of Salesforce
Salesforce signed an agreement to acquire Cresta, which runs voice agents for high-volume customer calls.
Partnerships & contracts 1
A large commercial agreement, or a government contract.
Dronamics sign a deal
- Value
- $100M
Dronamics now valued at $400M
Dronamics agreed to merge with the GP-Act III SPAC at a $400 million pre-money value, aiming to leave about $100 million in cash.
Moves
Who changed jobs in AI.
On the calendar
What's coming up, in date order, and how sure we are about each.
- 2–13NovProtoStar climate challengeToday
- 5–7NovThird LA Vertical Drama MarketToday
- 5–6NovBoxWorks 2026 in San FranciscoToday
- 6–7NovProtests call for a data-centre pauseToday
- 6NovOllie assistant shuts downToday
- 7NovTerry McMillan Presents: Paradise with You premieresTomorrow
- 9–20NovCOP31 climate conference in Antalya3days away
- 9–11NovThe Hinton Lectures on AI safety3days away
- 9NovCharmed By a Southern Monster premieres3days away
- 11NovJules becomes read-only5days away
- 12NovUpdated Claude usage policy takes effect6days away
- 15NovComments close on China's draft tech-export list9days away
- 15NovWorkBuddy and CodeBuddy free period ends9days away
- 18NovWebinar on running a dock with Stretch12days away
- 20NovAI Film Award submissions close2weeks away
- 20NovConductor cloud service shuts down2weeks away
- 20NovUpdated OpenAI usage policies take effect2weeks away
- 21NovInaugural edition in Hollywood2weeks away
- 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)2weeks away
- 29NovDaybreak Blue sessions can be used for training3weeks away
- 30–3Nov–DecContent London 20263weeks away
- 30NovRose d’Or Awards3weeks away
- 1DecHearing on Apple's bid to block testimony4weeks away
- 2–3DecWorkshop on autonomous surgical robots4weeks away
- 4DecAI-made film opens in limited release4weeks away
- 4DecGods Don’t Give Gifts premiere4weeks away
- 4DecSingapore AI Safety Fellowship ends4weeks away
- 31DecHy4 half-price offer ends8weeks away
- 31DecJules shuts down8weeks away
- Later in NovGuided Learning pilot in US schools
- Mid-NovXREAL Aura shipments begin
- H1 2027Digit 5 early customer access
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.