AI Daily Digest October 31, 2026: Grok Bot Gets a Wallet, Cloud Earnings Diverge, Flaws Hit Five Coding Tools

XAI announced that its Grok Bot agents can now complete purchases autonomously, using a rule engine called Auto Review with up to 20 Allow automatically and 20 Ask first rules, where Ask first takes precedence and payments via Vault-saved cards skip per-transaction approval. Amazon reported third-quarter revenue of $201.9 billion with AWS growing 39.3 percent to a $184 billion run rate and a $578 billion backlog, but capital expenditure of $77.3 billion drove trailing free cash flow to negative $42.4 billion and the stock fell more than 10 percent after hours. A security team also published working attacks against the browsers embedded in five coding tools.

The last day of October belongs to money and to mistakes. xAI gave its bots a payment rail, Amazon and Intel closed out a quarter that showed how differently the AI bill lands on each balance sheet, and a security team published working attacks against the browsers embedded in the tools many of us code with every day. Microsoft, meanwhile, quietly rebuilt the plumbing underneath Copilot so models can be swapped like parts. Here are the seven stories that mattered. xAI announced that Grok Bot, its team of always-on agents, can complete purchases for you. The pitch is short: connect Link, hand your Bot a shopping job, and it checks out on its own. That moves the product from an agent that drafts and researches to one that transacts, which is a different risk category entirely. The approval model is worth reading before you switch it on. 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 as options. You can set up to 20 Allow automatically rules and 20 Ask first rules, each up to 1,000 characters, and when both match, Ask first wins. xAI's own documentation lists purchases and financial transfers among the actions that should sit behind explicit boundaries, and it is candid that Auto Review is model-based, so it should complement those rules rather than replace them. Payment handling has the sharpest edge. Bots check out using a login saved in Vault, a connected password manager, or a Secure Form, and a card saved in Vault fills in at checkout without a separate approval each time. If you want to approve every payment, you have to leave the save-to-Vault box unchecked. Approval also does not undo completed work, so a wrong Allow is not a Ctrl+Z. This is the week's clearest sign that agent wallets are becoming a product category. Anthropic put virtual cards in Claude's hands earlier this month, and now xAI has a consumer-facing checkout flow with its own rule engine. The interesting part is no longer whether agents can spend money. It is who writes the rules they spend under. — xAI 🔗 xAI News https://x.ai/news Amazon's third quarter, announced Thursday, looked strong on almost every operating line. Revenue came in at $201.9 billion, up 12 percent, and operating income jumped 61 percent to $28.1 billion, a 13.9 percent operating margin that the company has never hit before. AWS grew 39.3 percent, its fastest pace in nearly five years and its sixth straight quarter of acceleration, putting the cloud business on a $184 billion annualized run rate. The AI demand story inside those numbers is real. AWS backlog reached $578 billion, Fortune 500 spending on Bedrock ran close to seven times last year's level, and token volume was up nearly seventeenfold. CEO Andy Jassy told analysts AWS should grow more than 40 percent in both 2027 and 2028, with a long-term path toward a trillion dollars in annual revenue. But capital expenditure landed at $77.3 billion for the quarter, up roughly 120 percent year over year and larger than the $42.4 billion of operating cash flow. Free cash flow over the trailing twelve months swung to negative $42.4 billion, long-term debt has climbed to $173.1 billion, and guidance for the fourth quarter came in below expectations. The stock fell more than 10 percent after hours. Net income slipped 2 percent, though that comparison is distorted by a $9.5 billion pre-tax gain on the Anthropic investment a year ago. Retail carried its weight, with North America operating margin at 7.7 percent, while advertising at $19.6 billion grew 11 percent but missed. Away from the spreadsheet, Amazon said Alexa for Shopping has reached more than 400 million customers in the past year, with users of its agentic shopping features up eightfold since the start of the year. Investors have now read three mega-cap AI quarters in one week, and the question they keep asking is the same one: when does the capex start paying them back. — Amazon · CNBC 🔗 Amazon News https://www.aboutamazon.com/news/company-news/amazon-news-today-top-stories-company · CNBC https://www.cnbc.com/2026/10/29/amazon-amzn-q3-2026-earnings-report.html Intel reported a quarter that shows exactly where the AI buildout is touching the older side of the chip business. Revenue was $16.5 billion, up 21 percent, and adjusted earnings of $0.52 a share beat expectations. The standout line was data center and AI revenue at $7.3 billion, up 78 percent, the fastest growth that segment has posted in fifteen years, driven by what CEO Lip-Bu Tan described as surging server CPU demand from AI agents. Then came the guidance. Intel projected fourth quarter revenue of $15 to $16 billion against analyst expectations near $17 billion, with adjusted earnings of just $0.25 a share. The stock dropped more than 20 percent after hours. Complicating the quarter, an explosion at the Kiryat Gat fab in Israel in October hit 10-nanometer server production, cutting an estimated $300 million from fourth quarter revenue and about two points of gross margin. Supply, not demand, is the binding constraint. Tan said global server CPU shipments should grow more than 30 percent in both 2026 and 2027 and that tightness could persist into 2028, while Intel 3 capacity and organic substrates for EMIB advanced packaging remain short. New capacity from an Oregon expansion is not expected until the second half of 2027. There were brighter spots elsewhere. AI PC related revenue rose 35 percent, the 18A-based Clearwater Forest server chip has started shipping, and foundry revenue grew 58 percent to $6.7 billion, though that unit still lost $1.9 billion. Intel enters 2027 with demand it cannot fully serve, a fab accident to digest, and a PC business that it expects to shrink 12 percent this year as memory prices bite. — Intel · Reuters Microsoft published a post on Tuesday that reads like a quiet architectural declaration. Written by Tina Schuchman, CVP of Foundry, and Vladimir Fedorov, CTO of GitHub, it introduces the Microsoft Copilot harness, the layer that lets one agent runtime direct work across many models instead of being married to a single one. The framing matters because model choice is becoming a routing decision rather than a procurement decision. Enterprises want frontier reasoning for hard planning steps and small fast models for the high-volume execution that dominates an agent's token budget, and they want to change their minds without rebuilding workflows. The harness post is the formalization of that idea across Foundry and GitHub. The clearest example ships alongside it. MAI-Code-1.1-Flash, Microsoft's small coding model, now runs in production inside GitHub Copilot at a quarter of the cost of the June original, with a 22 percent improvement on Terminal-Bench 2.1 in Copilot CLI, 15 percent better .NET task performance, and 25 percent fewer tokens per task. It can also read screenshots and UI mockups, which the June model could not. The more provocative piece is the local version. A 3-bit quantized build keeps the full 256K context window and matches the full-precision model on SWE-Bench Verified and Terminal-Bench 2.1, running with zero inference charges on hardware Microsoft recommends sizing above 120GB of RAM. Experimental access lands in the Copilot app, CLI, and VS Code by the end of October, with Auto routing deciding what stays on device and what goes to the cloud, and tool calls isolated in Microsoft Execution Containers. The coding assistant is quietly becoming a model portfolio managed by a harness, on your machine and Microsoft's. — Microsoft · microsoft.ai 🔗 Microsoft Blog https://blogs.microsoft.com/blog/2026/10/28/introducing-the-microsoft-copilot-harness/ · Microsoft AI https://microsoft.ai/news/mai-code-1-1-flash-br-better-faster-at-a-quarter-of-the-cost/ · MAI-Code repo https://github.com/microsoft/MAI-Code A security team called Forever Security published research this week, under the name EMBRACE, describing a shared root cause across five desktop AI assistants: Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop. Each embeds a browser that is trusted like a real browser while being given capabilities a normal browser does not have, and that combination breaks the browser security model the products were built on. The demonstrations are concrete. In Cursor, a web page could get code running on the host machine, shown in a demo that launches Calculator from a cold start with a single Enter. In GitHub Copilot, a page could spoof the address bar, painting github.com/login over an attacker's phishing page, which misleads the agent as much 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 at publication was mixed, with Cursor disclosed on September 10 and patched by October 1. The structural issue is less about any single bug and more about privilege: coding agents already hold tokens, credentials, and file access, and an embedded browser that blurs the trust boundary between web content and local capability turns every rendered page into potential attack surface. It lands in an odd week for AI security framing. Google is shipping its Gemini 4 Carbon model only to trusted cyber defenders, and Mistral pitched its newly opened trillion-parameter model partly on sovereign cyber defense, yet the demonstrated weaknesses sit inside the everyday tools on developers' desks. Supply is selling safety while the attack surface grows in the tools themselves. — Forever Security · The Next Web Neuralink showed off a striking upgrade this week. A patient identified as Kenneth, who has lived with ALS for seven years, can now trigger a Grok agent by mentally saying "Hey Neuralink" and thinking through the task he wants. The agent then takes over the computer screen and works autonomously, a demo of which made it onto YouTube. The change is architectural, not cosmetic. Earlier brain-computer interfaces translated neural activity into cursor movements, which meant a human supervised dozens or hundreds of small steps. Here the brain supplies one thing, an intention, and the agent handles the decomposition and the clicking. The interface became an input method for AI rather than a replacement mouse. A second demo may matter even more to the people involved. The implant converts intended speech into a clean audio stream injected in real time into the iPhone call stack, which is how Kenneth placed a phone call to his wife by thought alone, seven years after losing his voice. Neuralink engineers spent a 72-hour sprint living at the patient's home to ship eight new features. The limits are physical. Bandwidth, electrode count, and decoding speed all have ceilings, and no amount of agent cleverness gets around them. But the strategic read is hard to miss: the contest over where agents live started with browsers and chat windows, and it now includes the nervous system. Whoever owns the input layer owns the first move of every task. — Neuralink · 36Kr 🔗 36Kr https://www.36kr.com/p/4048035636558087 Anthropic announced an expanded agreement with the US General Services Administration that gives government partners 40 percent off Claude Enterprise, Claude for Government, and the Claude API, across every level of government. The deal moves Claude further into routine federal procurement rather than pilot programs. Context helps here. Earlier government AI deals were famously symbolic, with agencies enticed at around a dollar to experiment. A 40 percent discount on the commercial stack is a different posture: it assumes agencies will run real workloads and prices accordingly. Claude for Government has already cleared FedRAMP High, which removes the usual compliance detour for sensitive work. For Anthropic the timing is notable. The company is fresh off an S-1 filing that showed a business scaling fast on enterprise and developer demand, and public sector is one of the few customer categories where a discount can purchase distribution across thousands of agencies at once. Rivals have their own government channels, so the discount is also a wedge. The week's pattern across these stories is consistent. Whether it is Amazon absorbing capex, Intel rationing server chips, Microsoft routing models, or Anthropic pricing for Washington, the AI economy is moving from demonstrations to procurement, where discounts, supply constraints, and rule engines matter more than benchmark charts. — Anthropic KD Agentic · AI Daily Digest