● Grok Bot can shop xAI says the Bot can now complete purchases on your behalf once you connect Link.
● Gemini 4 Carbon Number one on the Artificial Analysis Intelligence Index v5.0, scoring 59. Cyber defenders only, for now.
● Mistral Large 4 weights Trillion-parameter open weights on Hugging Face, pitched at sovereign cyber defence.
Alphabet, Microsoft and Cohere numbers below, plus the research that broke five AI coding browsers.
Two threads ran through the 48 hours to 30 October. The first is agents being handed a wallet. The second is that three separate announcements all pitched the same job - cyber defence - in the same week that researchers published working attacks against the AI coding tools developers already use.
1. Grok Bot can now buy things on the internet
xAI announced that Grok Bot can complete purchases on a user’s behalf. The company’s own post puts it plainly: connect Link, send your Bot shopping, and it can complete purchases for you. This moves Grok Bot from an agent that drafts and researches to one that transacts.
The approval model is documented and worth reading before you turn it on. xAI’s own documentation lists purchases and financial transfers among the actions that should sit behind explicit boundaries. Every action runs through Auto Review, and anything matching an Ask first rule produces an approval card showing the proposed operation and its inputs, with Allow once, Always allow or Deny. You get up to 20 Allow automatically rules and 20 Ask first rules, each up to 1,000 characters, and Ask first wins when both match.
The detail that matters most is in the payment handling. Bots check out using a login saved in Vault, a connected password manager, or a Secure Form - and xAI states that a card saved in Vault fills in at checkout without a separate approval each time. If you want to approve every payment, leave "Securely save to Bot’s vault" unchecked, or remove the card under Settings → Vault. xAI also notes that approval does not undo completed work and that Auto Review is model-based, so it should complement explicit rules rather than replace them.
2. Gemini 4 Carbon takes the top of the Artificial Analysis index
Google launched Gemini 4 Carbon on 28 October, describing it as its strongest model with state-of-the-art cybersecurity and long-horizon capabilities. It ranks number one on the Artificial Analysis Intelligence Index v5.0 with a score of 59.
The security benchmarks are the point of the release. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation which it calls a new state of the art, and 71.0% on CWE-bench for patching. On the Gray Swan indirect prompt injection benchmark it reports a 0.0% attack success rate against 0.7% for Argon. On WebDev Arena it sits fourth at 1768 Elo, up 89 points.
Pricing is introductory at $2 per million input tokens and $10 per million output, with cached input 95% off, rising to $4 and $20 once the introductory period ends. Availability is the catch: Carbon is not broadly released. It is going to a subset of cyber defenders first, expanding to trusted partners through the Fairwind Program, with developers and consumers promised later.
3. Mistral opens the weights to its trillion-parameter Large 4
On 29 October Mistral released the weights for Mistral Large 4, the trillion-parameter model it nicknames Le Chonk, on Hugging Face and through its own API. The pitch is sovereign cyber defence for enterprises and governments: detect, analyse and respond to threats across code and IT infrastructure, deployed on-premises if you want it there.
Pierre Stock, Mistral’s VP of Science, made the argument for open weights directly - access "can’t be cut off by definition." That is the sharpest contrast with the Gemini 4 Carbon release above, which ships the same capability to an approved list.
Mistral claims Large 4 is comparable to GLM-5.3, widely considered one of the strongest open-weight models for cyber work, while using 47% fewer tokens per task. It also claims the model outranks all other open models on Artificial Analysis’ Coding Agent Index.
4. Researchers broke the browsers inside five AI coding tools
Forever Security published EMBRACE, research into a shared root cause across five desktop AI assistants: an embedded browser trusted like a real one. The browsers built into these tools were given capabilities a normal browser does not have, which broke the browser security model.
The affected products are Cursor, GitHub Copilot, Codex, VS Code and Claude Desktop. In Cursor, a web page could get code running on the host - the researchers demonstrate launching Calculator from a cold start with one Enter. In GitHub Copilot, a page could spoof the address bar, showing github.com/login over an attacker’s phishing page, which misleads the agent as well as the human. In Codex, Claude Desktop and VS Code, a local file could escape its project folder and read files across the disk.
Patch status as published: Cursor was disclosed on 10 September and patched on 1 October. GitHub fixed the Copilot address-bar spoof independently in version 1.1.29 on 13 October, before the researchers’ report arrived. Codex and VS Code remain unpatched - Codex was disclosed to OpenAI on 27 July with no response, and VS Code was disclosed to Microsoft on 10 September and declined.
5. Alphabet Q3: revenue up 25%, Cloud up 103%
Alphabet reported third-quarter revenue growth of 25% year over year, to roughly $128 billion, with Google Cloud revenue up 103%.
The two usage figures are the ones worth keeping. Alphabet says its model APIs are now running at nearly 32 billion tokens per minute, 40% higher than last quarter. And the Gemini app has reached one billion users.
6. Microsoft Q1 FY27: Azure up 48%, backlog up 73%
Microsoft reported first-quarter revenue of $92.3 billion, up 19%, with Microsoft Cloud at $63.9 billion, up 30% (31% in constant currency) and Azure revenue up 48%. Operating income was $45.9 billion, up 21%.
The number that says most about the next two years is the backlog: commercial remaining performance obligation increased 73% to $677 billion. That is contracted revenue not yet recognised, and it is the clearest signal in the release of how much capacity customers have already committed to buying.
7. Cohere closes the first tranche of a Series E, with Canada investing
Cohere announced on 29 October that it has completed the first close of a Series E it aims to grow to as much as C$7 billion, at a valuation the company puts at C$35.7 billion. Bloomberg reports the round could total up to $5 billion at a $20.5 billion valuation excluding the new money.
The notable investor is a government. The Government of Canada has agreed to invest up to $1.14 billion, per Bloomberg, joining Nvidia, Inovia Capital and Prosperity7 Ventures from earlier rounds. Schwarz Group companies lead the round, having committed $600 million earlier, with Bell Ventures and Bullhound Capital also new. Cohere’s CFO said the first close includes "most of it" without putting a number on the committed share.
8. DOE commits $5 billion to nuclear uprates that feed AI data centres
The Department of Energy is providing $5 billion in loans to Constellation Energy to expand nuclear generating capacity at eight power plants by 1.1 gigawatts. The commitment is conditional and not yet finalised.
The plants are Braidwood, Byron, Clinton, Dresden, La Salle and Quad Cities in Illinois, Calvert Cliffs in Maryland and Limerick in Pennsylvania. The buyers are already named: Google takes additional output from the Pennsylvania plant and four of the Illinois plants, Amazon buys from Calvert Cliffs, and Meta effectively takes the entire Clinton plant. Constellation says it has committed more than $7 billion to add 1,080 megawatts for those three.
9. Anthropic expands AI for Science to 20 institutions
Anthropic expanded its AI for Science programme on 28 October, adding partnerships with 20 scientific foundations, institutions and research groups worldwide.
What the week adds up to
Put the releases side by side and the pattern is hard to miss. Google shipped a frontier model whose headline benchmarks are vulnerability discovery, exploitation and patching. Mistral shipped a trillion-parameter open-weight model pitched at exactly the same job. Both landed within 48 hours of research showing that the browsers inside five widely used AI coding tools could be made to run code, spoof a login page and read files across a disk - with two of the five still unpatched.
| Shipped | Cyber claim | Who can use it |
|---|
| Gemini 4 Carbon | #1 on Artificial Analysis at 59; 88.2% SEC-bench Pro | Subset of cyber defenders, via the Fairwind Program |
| Mistral Large 4 | Sovereign cyber defence; 47% fewer tokens than GLM-5.3 | Anyone - open weights on Hugging Face |
| EMBRACE research | Code execution, address-bar spoofing, file theft | Codex and VS Code unpatched at publication |
The uncomfortable part is the sequencing. The industry is selling defensive capability at frontier prices while the development tools it sells to the same buyers carry unfixed flaws disclosed three months ago. Grok Bot gaining a payment method in the same week is not unrelated: every capability added to an agent is also an increase in what a successful prompt injection is worth.
If you use any of the five tools named above, the practical step today is checking your version rather than reading another benchmark table. Cursor is patched as of 1 October, Copilot as of 1.1.29 on 13 October.
Related reading
Sources
Every figure on this page was read from the source listed above on 30 October 2026. Where a company and a news outlet give different numbers for the same round, both are shown rather than averaged.