Daily Briefing
23 October 2026 · 5 stories
🔥 Top story
1Anthropic hands its agent recipes to everyone
On 22 October, Anthropic handed a neutral foundation, created under the Linux Foundation, the recipe format for our agents. A recipe, in their language, is a simple text file called SKILL.md: you write down one task, and the agent reads it when the time comes. It is the instruction manual you put in a drawer rather than rewriting from scratch every time something breaks. The format is already understood by more than seventy platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code. Anthropic reports one million recipes published and close to 280 million installations. On the same day, OpenAI, Hugging Face and Alibaba deposited their own projects in that foundation, including an Alibaba isolation box used to keep agents contained. A standard stays a standard even when the people who invented it stop controlling it. Source : Agentic AI Foundation.
2Pirates used an agent against 2,250 companies
Microsoft accuses a Chinese group called Storm-0940 of using ARTEX, a free AI-driven penetration testing tool, to automate attacks against more than 2,250 companies that were customers of a backup provider. The entry point was not even a stolen password: it was the access the provider had already obtained for its customers, reused without a single login. Between 10 and 14 October, more than 151,000 attempts failed. On 15 October, the day the flaw was published, the method changed: the agents logged in with the provider credentials at a rate of 2,900 attempts per minute, and went from the initial target to privileged access in under three hours. ARTEX is free, public, and works with any model. For a company, the lesson is simple: access granted to a supplier is a master key to your whole estate. Source : Microsoft Security Blog.
3Walmart opens its tills to agents, Amazon blocks them
The two largest American retailers do not agree on what to do with robots that do the shopping. Walmart opens the door: every AI agent gets a customer profile of its own, with a spending limit and a confirmation before each purchase, exactly like when you add someone you trust to a bank card. The agent cannot see payment details or purchase history, and the returns policy stays the store's. Walmart even issues refunds when the agent gets it wrong. Amazon does the opposite and has been blocking shopping agents since 20 September. The real problem is identity: according to an analysis published on 20 October, only 2.3 per cent of legitimate agent sessions carried a verifiable signature. How do you know who is speaking? Source : TechCrunch.
4Claude Duet keeps working with your computer off
Anthropic unveils Claude Duet, an agent that does not wait for your questions to make progress. Every morning it prepares a summary, it keeps working between two of your messages, and it only asks your permission when a decision genuinely deserves it. It runs on its own virtual computer, with its own browser: you can turn your machine off or walk away, and it carries on. It connects to more than 6,000 applications, from Gmail to Shopify, and can pay for purchases. Two details matter more than the rest. First, sensitive actions require approval, but transferring money is forbidden, not merely supervised. Second, Anthropic admits the first version may misjudge how urgent a task is. That is exactly the line we follow in our own safeguards. Source : The Verge.
5Beijing puts model weights under licence
China is preparing an identity card for the most powerful artificial intelligence models. The draft rule, published on 16 October, makes government authorisation mandatory to export models that exceed ten to the twenty-sixth power in computing operations, two trillion parameters, or that sit in the world's top ten. The weights, the source code, the training methods and even AI agents are all covered. But here is the exception that changes everything: serving a Chinese model abroad through a website or an automatic connection stays licence-free. In other words, you cannot take the model away, but you can use it remotely. The public consultation closes on 15 November. Source : New York Times.
📡 To watch
DeepSeek raises 15 billion dollars
The Hangzhou lab has raised 100 billion yuan, the largest Chinese round of the year, with battery maker CATL and tech giant Tencent as lead investors. The amount was tripled along the way, and DeepSeek is preparing an initial public offering. The detail that matters: a few days earlier, the same models had been excluded from United States federal civilian systems. The money and the ban arrive in the same envelope. Source : Straits Times.
OpenAI targets 1.25 trillion before its listing
OpenAI is said to be closing a round of up to 30 billion dollars at a valuation of 1.25 trillion before the money comes in, led by United Arab Emirates investors and the Coatue fund. A month ago the figure was 1.4 trillion: the gap shows the terms are still moving. Management puts the listing in 2027. Anthropic is said to be aiming for 1.5 to 2 trillion for its own. Source : TechCrunch.
One test, 0.22 dollars against 2.43 dollars
On a 3D animation, a Chinese model delivered the result for 0.22 dollars, against 2.43 dollars for the leading American model, and 2.3 times faster. Except the Chinese model forgot the last step of the request. Cheapest is not the same as most reliable. Source : The Hype.
The same hacking tool now targets banks
ARTEX is not reserved for Storm-0940. Zscaler and Recorded Future have seen a Sino-Korean group use it against South Korean financial institutions, with DeepSeek, GLM and Groq as coordinated agents, one checking the work of another. A free open-source tool becomes a weapon on both sides. Source : Microsoft Security Blog.
📊 Trend
Three threads cross today. First, the agent doorstep has become a question of power. Recipe standards are handed to everyone, agents are let through the till with a spending cap, yet they are forbidden from roaming loose inside a Chinese model, and they are given a sanitised server to break into a single flaw: everywhere the question is no longer what an agent can do, but how far it is allowed to go. Second, the identity of the agent is the battle. One retailer gives every shopping bot a name and caps its spending; another blocks it; and a study published on 20 October shows fewer than 3 per cent of legitimate sessions carried a verifiable signature. Without reliable identity, every authorisation rests on sand. Third, price no longer decides alone. A Chinese model delivered an animation for 0.22 dollars against 2.43 dollars and forgot the final step; an American model delivered the complete loop, even though it costs eleven times more. In our own experience the same reflection produced the same conclusion last month, when an incomplete brief made the deliverable useless. What decides is not the price: it is completeness. Source : The Agent Watch.