1 The news2 For your work3 Money4 Before you goGemini 4 Carbon Image: Google Google Google's new frontier model tops an independent intelligence ranking and leads on vulnerability-finding tests, but at first only selected cyber defenders can use it.
Google DeepMind has announced Gemini 4 Carbon less than a week after releasing Gemini 4 Argon. It describes Carbon as its strongest model for cybersecurity and long-horizon work. The independent evaluator Artificial Analysis ranks it first on its Intelligence Index v5.0 with 59, one point ahead of Anthropic's Claude Opus 5.5 and two ahead of OpenAI's GPT-6 Astra. Google reports leads on vulnerability benchmarks, including 88.2% on SEC-bench Pro, and says its own teams have run the model for days on multi-step engineering tasks. At first, access is limited to selected cyber defenders. The introductory price matches Argon at $2 per million input tokens and $10 per million output tokens. A footnote says this later rises to $4 and $20, with no end date given.
Introductory price $2 input / $10 output per million tokens
Later price $4 / $20
Access Initially for select cyber defenders
SEC-bench Pro 88.2% Why it matters Frontier models that find and exploit vulnerabilities change both cyber defence and the risk of attack, and vendors are now rationing access.
Artificial Analysis Intelligence Index v5.0 Model Score Gemini 4 Carbon 59 GPT-6 Astra 57 Gemini 4 Argon 50
Widely reported The Verge Ars Technica The Hacker News Artificial Analysis
Part 1 of 4 The news What changed in AI today, lab by lab, and who answers for it.
OpenAI has updated its Model Spec, the public document describing how its models should behave, reorganising the guidance around five values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. The company says the aim is for models to apply human values in unfamiliar situations.
Published October 29, 2026 Why it matters The Spec sets out what organisations building on OpenAI's models can expect them to refuse, prioritise and disclose.
Anthropic has named 20 foundations, institutions and research groups it has partnered with as it expands its AI for Science programme, which gives free Claude credits to researchers on high-impact projects. Partners include the Sloan Foundation and San Francisco's Gladstone Institutes; the programme previously supported individual scientists.
Why it matters Labs are competing for scientific users through credits and partnerships, which may open new routes to subsidised model access for research bodies.
Image: The Next Web Alphabet shares fell more than 12% after the company said searches in AI Mode earn less than traditional ones, wiping $511bn from its value. Search revenue growth slowed to 12%. Google is responding by taking AI Mode ads to 10 more countries, with Brazil, Mexico and the UK next.
Share fall More than 12%
Search revenue growth 12%, from 19%
AI Mode ads 10 more countries Why it matters AI answers are changing how search makes money, which affects advertisers' costs, publisher traffic and how businesses are found online.
Image: TechCrunch Meta has released Style Sketch, an AI fashion app that connects to Instagram to build a virtual wardrobe from photos users have already shared. Items can also be added by photographing them, recording a short video or uploading images from the camera roll, without a formal launch announcement.
Why it matters It shows Meta testing standalone AI consumer apps built on Instagram data, a pattern that brands and retailers will watch.
Amazon Web Services revenue rose 39% to $46 billion in the third quarter, its fastest growth in nearly five years, and Andy Jassy expects it to exceed 40% in 2027 and 2028. Amazon's earnings missed estimates and its shares fell more than 7% after hours.
AWS Q3 revenue $46 billion
AWS run rate $184 billion
Amazon revenue $201.9 billion Why it matters Cloud capacity for AI is still growing fast, and providers' heavy spending shapes prices and availability for their customers.
Microsoft has disclosed Azure's quarterly revenue in dollars for the first time: $33.2 billion, up 48%, beating expectations amid demand for cloud and AI. The disclosure allows a direct comparison with Amazon Web Services, the largest cloud provider, which has long reported its revenue in dollars.
Azure growth 48% Why it matters Clearer cloud figures help buyers judge the scale and momentum of the providers they depend on for AI capacity.
Image: TechCrunch Other labs · Thrive Holdings Thrive Holdings, a permanent-capital company backed by OpenAI, has agreed to buy the home-security firm ADT in an all-cash deal at an enterprise value of about $15.3 billion, a 45.6% premium. It plans to bring AI to the business. ADT may seek rival offers until 28 November.
Enterprise value $15.3 billion
Equity value $7.6 billion
Premium 45.6%
Rival-bid window ends November 28, 2026 Why it matters It tests whether AI-focused owners can lift margins in established service businesses with large customer bases.
Image: Mistral AI Mistral has published the weights of Mistral Large 4, nicknamed Le Chonk, on Hugging Face after three more weeks of reinforcement learning. The 1 trillion-parameter model activates 52 billion parameters per task. Mistral says it is the best open-weight model for coding, with strong cyber-defence ability.
Size 1 trillion parameters, 52 billion active
Runs on Quantised version on four Nvidia B200 or B300 GPUs
Licence Mistral Ecosystem Open-Weight License (modified MIT)
Terminal-Bench 4.0 41%, from 28% Why it matters Organisations that want a frontier-class model they can host and control themselves now have a European option, if they have datacentre hardware.
03 Rules, safety and accountability Regulation, sovereignty, safety, ethics and liability.
OpenAI has described a multilayer system that watches its models for misaligned behaviour across training, evaluation, internal use and external deployment. Cheap probes screen traffic, and a random 5 percent of unflagged cases goes on for deeper review. OpenAI says it paused frontier reinforcement learning while it strengthened agent containment.
Coverage Training, evaluation, internal use, external deployment
Random review sample 5 percent of unflagged traffic Image: Interesting Engineering What to do Log agent actions somewhere the agent cannot reach, sample unflagged activity for human review, and alert whenever monitoring settings change.
Image: Ars Technica Sovereignty · US government President Trump has signed a Section 232 proclamation finding that imports of robotics, industrial machinery and their parts threaten US national security. It imposes no tariffs for now. Instead it directs the Commerce Secretary and the Trade Representative to negotiate agreements with trading partners, with further action possible later.
Why it matters Buyers of robotics and automation face possible tariffs and supply changes as the US treats machinery as strategic.
Image: ChatGPT Is Eating the World Liability · AI developers The litigation tracker ChatGPT Is Eating the World counts 157 US copyright suits against AI companies after adding 9 new cases. Separately, the YouTube creator Ted Entertainment has voluntarily dropped its suits against OpenAI, NVIDIA, Apple, Amazon, ByteDance and Meta, though the co-plaintiff Matt Fisher remains.
Total US suits 157
Newly added 9 Why it matters Copyright exposure remains a live cost for AI developers, and the outcomes may shape the indemnities and licence terms offered to customers.
Residents near xAI's data centres in Mississippi say the company has offered to buy their homes on condition that owners do not join a class-action nuisance suit over gas-turbine noise. Forbes, citing the Guardian, reports that the offers target 418 homes, almost all of the Colonial Hills neighbourhood.
Why it matters Local opposition to noise, power and water use is becoming a cost and a planning risk for data-centre projects.
Reward · Victorian Legal Services Board + Commissioner A census of 2,221 Victorian lawyers by the state's Legal Services Research Centre found that 55% use AI in their practice, up from 37% in 2025. Among AI users in fee-charging practice, 26% said AI had led to reduced fees for clients, and most reported no change to pricing.
Using AI 55%
In 2025 37%
Respondents 2,221
Reduced fees 26% of AI users Why it matters It is measured evidence that AI adoption in professional services does not automatically pass savings on to clients.
Surge AI has released DAYJOB: Tax, a benchmark that asks AI agents to complete 50 realistic tax assignments end to end, from a client request to a finished deliverable, working through messy records and exercising professional judgment. A public leaderboard ranks the models evaluated, and the dataset is on Hugging Face.
Assignments 50 Image: Surge AI Why it matters Benchmarks built on whole professional tasks give firms a better guide to agent readiness than short question sets.
Part 2 of 4 For your work Tools to use today, one thing to try, security to act on, and how work is changing.
GitHub's Copilot Fusion, now generally available, picks and combines models for each coding task instead of leaving the choice to the developer. It offers Standard, High and Max effort levels, uses only models an organisation's policies allow, and bills each model used at its standard rate.
Use it Select Fusion as the model in VS Code or Copilot CLI and compare cost and results over a week.
Runs in VS Code, Copilot CLI, Copilot app, JetBrains IDEs
Effort levels Standard, High, Max
Billing Each model used at its standard rate
Admin control Only models enabled in model policies Image: GitHub Why it matters GitHub says routing lowers cost, but spend now depends on models the tool picks, not the developer.
Through Amazon Bedrock AgentCore Gateway, agents can now call AWS Step Functions workflows as tools, deciding when to run one and what input to pass. A built-in connector generates the tool schemas and authenticates with the gateway's service role, while Step Functions runs the workflow itself.
Use it Expose a refund or onboarding workflow to an agent so it runs the approved steps every time.
Connector ID states
Tool discovery Agent calls tools/list on the gateway
Authentication Gateway service role Why it matters Agents can trigger audited business processes without teams rewriting them as new code or exposing raw APIs.
Cloudflare's Agents SDK now treats Agent Skills as files: a folder containing a SKILL.md can live in an agent's Workspace, an R2 bucket or code. Each SKILL.md is checked against the open Agent Skills specification, and scripts in a skill can run through Workspace exec.
Use it Move your team's existing SKILL.md folders into an agent's Workspace instead of pasting them into prompts.
Skill sources Workspace, R2 or code
Format SKILL.md, checked against the Agent Skills specification Why it matters Skills written once for coding agents can be reused inside production agents, cutting duplicate prompt maintenance.
Cloudflare Add Cloudflare's official Agent Skills to Claude Code Cloudflare publishes a set of Agent Skills for building on Workers and the Agents SDK, installable as a Claude Code plugin. They follow the open Agent Skills standard, so the same folders also work in other compatible agents.
Step 1
Add Cloudflare's skills marketplace inside Claude Code
/plugin marketplace add cloudflare/skillsCopy Step 2
Install the Cloudflare plugin from that marketplace
/plugin install cloudflare@cloudflareCopy A campaign planting cryptominers and backdoors through Hugging Face model repositories reappeared within days of removal, Pluto Security found. It abuses the trust_remote_code option, which lets a model ship and run its own Python code, and one model was downloaded over 28,000 times.
Downloads Over 28,000 for one model
Method trust_remote_code
Found by Pluto Security What to do Keep trust_remote_code off by default and read the code of any model that needs it before loading. Check GPU servers for unexpected processes.
A researcher scanning 7.25 million public Gitee repositories found 26,400 credentials that still worked, mostly keys for cloud and AI services, led by 9,473 Alibaba Cloud AccessKey pairs, alongside DeepSeek and OpenAI keys. Vendors with the most exposed keys were told; the developers who leaked them were not.
Repositories scanned 7.25 million
Alibaba Cloud key pairs 9,473
Committed in 2025 12,242 What to do Search your repositories for committed keys and rotate any you find; deleting them from code is not enough.
Check Point has tracked a campaign, dubbed Yoba, that uses fake GitHub projects posing as desktop apps such as 'GPT-6 Astra' and 'Claude Opus'. The malware sits beside genuine signed OpenAI and Anthropic executables so it loads indirectly, ending with a cryptocurrency miner.
Delivered as ZIP releases on GitHub
Technique DLL side-loading beside signed executables What to do Install AI desktop apps only from the vendor's own website or official store, and treat ZIP releases from unfamiliar repositories as untrusted.
Security firm Dream says a suspected China-linked actor exploited PHP2Shell flaws and a Linux kernel bug to root 25 servers, three of them in a US government-affiliated network. The attacker targeted firewall management interfaces and used tools including Claude Code and DeepSeek to automate scanning and credential theft.
Servers rooted 25
Entry point Firewall management interfaces What to do Patch PHP and the Linux kernel on internet-facing systems, and keep firewall management interfaces off the public internet.
404 Media, citing eight current and one former employee, reports that Proton's chief executive is pushing staff to build with AI as fast as possible, and that the company will re-interview staff about AI use, interviews that could decide jobs. Proton did not dispute the reporting.
Why it matters It shows how mandated AI adoption, tied to job security, can collide with staff concerns over security and quality.
Law firm Ropes & Gray is developing a transactional diligence tool with OpenAI that takes a first pass through deal documents, grouping them and drafting summaries and reports. Lawyers review that output before advising clients, aiming to bring their judgement in earlier on time-pressed deals.
Why it matters Document review by junior lawyers is among the first legal work being restructured, with humans kept as reviewers.
10 How companies work with agents Engineering account MoEngage turned roughly 400 private APIs into agent tools through one gateway. In an account published on the AWS blog, MoEngage describes exposing roughly 400 private REST APIs across 7 AWS Regions as permission-scoped agent tools through Amazon Bedrock AgentCore Gateway. It says this avoided a separate MCP server per service and kept the endpoints off the public internet.
Inventory the APIs. List which internal services agents genuinely need before exposing anything as a tool.Scope permissions. Give each tool only the rights the calling user or agent should have.Centralise access. Route tools through one gateway rather than building and securing a server per service.Keep endpoints private. Make sure internal APIs stay unreachable from the public internet when agents call them.Part 3 of 4 Money Where AI’s money is going: rounds, valuations, listings, public funding and prices.
A Crunchbase News analysis argues that large AI rounds are hiding a shrinking Series A market. Seed-stage founders are receiving bigger cheques but face higher revenue expectations and longer waits before a Series A, according to the piece, which includes advice from an investor at MGV.
Why it matters For founders, the gap means AI headline rounds say little about how hard a typical Series A has become.
Funding rounds Company Stage Amount Valuation Led by Wabi Consumer AI assistant that works inside shared group threads Seed $55M $1.25B IVP Avante Enterprise AI platform for employee benefits teams Series A $30M – Frist Cressey Ventures Dream AI world models for construction trained on building data – $17M – a16z Plato Simulated workplaces where AI models train on real jobs Series A $16M – Nava Ventures Simplismart Software that helps companies run AI models more efficiently Series B $14M – Dallas Venture Capital Bleu Robotics AI models teaching industrial humanoid robots tasks from fewer demonstrations Pre-Seed €5M – – Pixel Systems Technology that identifies AI-generated material Seed €2.7M – – VUSE XR AI-powered mixed reality training software for enterprises Pre-Seed €1.2M – IACD
Listings
No AI company filed or listed today.
What models cost, per million tokens
No model changed its price today.
Public money, open now No new public funding found today.
Venture capital, in plain words
Pre-money valuation What investors agree a company is worth immediately before their new money goes in.
In practice It sets the price per share, and with the size of the round, how much of the company the new investors own.
Part 4 of 4 Before you go One long read, the rest of the day in a line each, and what is coming.
14 One long read Worth twenty minutes.
JiTTests Engineering at Meta · testing Meta's engineers explain how Catching Just-in-Time Tests are generated for each code change. They describe how a language-model judge ranks the results, and how a human-on-the-loop step means engineers review outcomes in aggregate, not every test.
Why it matters For engineering leads weighing AI-generated tests, and anyone who wants human oversight without engineers reading every machine-written output.
OpenAI now lets eligible ChatGPT Pro and Business subscribers claim up to $500 a month in API credits. OpenAI Developer Community GitHub's 2026 Octoverse report says AI agents started more than 40% of pull requests opened in September. Visual Studio Magazine Meta has described the next version of Andromeda, its ads retrieval engine, built by designing the model, system and GPUs together. Engineering at Meta ByteByteGo explains how Uber built a central gateway that lets its AI agents find and call internal tools. ByteByteGo An essay on Martin Fowler's site argues that experienced developers distrust coding agents because code that builds may still fail in production. Martin Fowler 2–4 Nov
Conference · Brooklyn A developer conference on building software with AI agents, held in Brooklyn with an online option.
16–20 Nov
Conference · Berlin A machine learning conference for practitioners, held in Berlin with an online option.
16–20 Nov
Conference · Berlin A summit in Berlin on the engineering practices behind production AI systems.
25–26 Nov
Conference · London A two-day conference in London on applying AI in business and public services.
30 Nov – 4 Dec
Conference · Las Vegas Amazon Web Services' yearly cloud conference, where the company usually announces new AI and infrastructure services.
