A Mystery Outage Snarls CSX's Freight Network, Belgium's Tax Portal Goes Dark After a Cyberattack, and Google's Gemini 4 Carbon Takes the Top Spot
By Noah Smith, Owner & Consultant, KeyChange Technologies · October 29, 2026
Here is your AI and IT news for Thursday, October 29, 2026: one of America's biggest freight railroads spent a second day running on paperwork, Belgium's tax portal went dark after a cyberattack, and Google shipped a model that tops the leaderboards but that you cannot use yet. Yesterday's recap is here if you missed it.
📌 The AI and IT news at a glance
- A technology outage slows CSX, one of the largest US freight railroads, and forces crews back to paper.
- Google DeepMind's Gemini 4 Carbon takes the #1 spot on a major AI leaderboard, but only cyber defenders get it for now.
- OpenAI reportedly plans to buy Google's AI chips, not just rent them.
- A cyberattack knocks Belgium's tax portal offline, and filing deadlines are being pushed back.
- Anthropic says an unreleased Claude model found and chained real flaws in Chrome, Safari, Android, Windows, and Linux.
- Hackers have quietly planted code in hundreds of Adobe Commerce and Magento online stores.
- Anthropic gives Claude Enterprise customers reports on what their AI spending actually produces.
- GitHub's Copilot Fusion goes live: pick an effort level and let Copilot pick the models.
🔝 Top story
A technology outage slows CSX and sends crews back to paperwork
CSX, the freight railroad that moves a big share of the goods crossing the eastern United States, spent Tuesday and Wednesday fighting a technology outage that slowed its operations. According to News4JAX in Jacksonville, where CSX is headquartered, key functions remained limited on Wednesday. Calling crews to trains required significant manual work, and CSX could not complete normal interchanges (the handoff of rail cars) with the other big Class I railroads. A limited interchange continued with the Belt Railway of Chicago and Conrail, using paperwork and manual processes.
The good news is what still worked. Train dispatching and positive train control, the safety system that can stop a train automatically, kept functioning. A CSX spokesperson said its security team confirmed it was "not a cyberattack and that no data breach occurred," and that systems were coming back online on a rolling basis. CSX has not said what caused the outage or when everything will be fully restored, and it says it is in regular contact with other railroads and federal rail regulators.
In short: A non-cyber technology outage slowed CSX's freight operations and forced manual crew calling and paper-based handoffs with other railroads.
What it means for your business: If you ship or receive goods by rail in the eastern US, expect some delays to ripple through this week. More broadly, it is a reminder that a big outage does not need a hacker; plain system failures can stop a business just as hard.
My take: The headline here is not "railroad has computer trouble." It is that CSX could keep its trains safe and moving at all because it still had people who knew how to do the job on paper. Most small businesses have no paper fallback for anything: not invoicing, not scheduling, not dispatch. Pick your one most critical process and ask what you would do for two days without the software. If the answer is "nothing," write a one-page manual version now while it is calm.
Source: News4JAX: Technology outage slows CSX operations; security team rules out cyberattack
🤖 AI
Google's Gemini 4 Carbon takes the top spot, but cyber defenders get it first
Google DeepMind announced Gemini 4 Carbon on Wednesday, calling it its strongest model yet for cybersecurity and long, multi-step work. It arrives just four weeks after Gemini 4 Argon. Google says Carbon ranks first on the independent Artificial Analysis Intelligence Index with a score of 59, a point ahead of Anthropic's Claude Opus 5.5 at 58 and OpenAI's GPT-6 Astra at 57. Google also reports state-of-the-art scores on tests for finding software vulnerabilities (88.2% on SEC-bench Pro) and patching them (71.0% on CWE-bench), and claims more than double Argon's rate of writing secure code.
The catch is access. Carbon is not generally available. Google is starting with a subset of cyber defenders this week, then expanding to trusted partners through what it calls the Fairwind Program, and it is taking part in the US government's voluntary pre-release testing. Developers, businesses, and consumers will get it "as soon as possible," with no date given. When it does open up, introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after an unspecified introductory period.
In short: Google unveiled Gemini 4 Carbon, now the top-ranked model on a leading independent index, but is limiting it to vetted cyber defenders for now.
What it means for your business: Nothing changes in your tools today. But the top three AI labs are now within two points of each other, which means no single vendor has a lasting lead, and you should avoid locking yourself into one.
My take: The pattern is now unmistakable: Google, Anthropic, and OpenAI are all holding their strongest cyber-capable models back for defenders first. That is the right call, and it tells you how seriously the labs take the attack risk. For a business owner, the practical move is boring: keep your AI tools swappable, and do not rebuild your workflows every time a leaderboard flips. It will flip again next month.
Source: Google: Introducing Gemini 4 Carbon
OpenAI reportedly plans to buy Google's AI chips
OpenAI plans to buy Google's custom AI chips, called TPUs, according to a Reuters report on Wednesday citing The Information. OpenAI already rents TPU capacity through Google Cloud, but a purchase would put Google's chips alongside the Nvidia GPUs that power most of OpenAI's work today. Financial terms and volumes were not disclosed, and neither company has confirmed a deal.
If it happens, it is a notable shift. Nvidia has had close to a lock on top-end AI chips, and the companies building AI have been scrambling for any capacity they can get. A rival lab buying hardware from Google, which competes with OpenAI directly through Gemini, shows how tight that supply still is.
In short: OpenAI reportedly plans to buy Google TPUs to add to its Nvidia-based computing supply; neither company has confirmed it.
What it means for your business: Chip shortages are a big reason AI prices and usage limits move around. More suppliers in the mix should, over time, mean steadier capacity and pricing for the AI tools you pay for.
My take: Treat this as a report, not a done deal. But the direction makes sense: when your biggest competitor has the chips you need, you buy them anyway. For the rest of us, more competition in AI hardware is good news for prices, even if it takes a while to show up on the bill.
Source: Reuters: OpenAI plans to buy Google's AI chips, The Information reports
🛡️ IT and security
Why these made the cut: a government tax portal that businesses depend on, AI finding real flaws in the browsers and phones almost everyone uses, and an ongoing attack on the software behind a huge number of small online stores.
A cyberattack knocks Belgium's tax portal offline
Belgium's Federal Public Service Finance announced on Wednesday that it had been hit by a cyberattack. Its online tax portal, MyMinfin, where individuals and businesses manage tax returns and financial documents, was taken offline, and several other online services were shut down as a precaution while officials investigate and secure their systems.
The agency says that so far there are no signs that anyone's personal data was breached, though the investigation is ongoing. It has not said who was behind the attack or how they got in, and it has given no timeline for restoring service. It will extend some upcoming tax filing deadlines affected by the disruption, with new dates to be announced.
In short: A cyberattack forced Belgium's finance ministry to take its MyMinfin tax portal and other online services offline, and some filing deadlines will be extended.
What it means for your business: When a government portal you file through goes down, the deadline is still yours until the agency officially moves it. Keep proof that you tried to file, and do not leave filings to the last day.
My take: This comes the same week the US Federal Election Commission's filing system went dark. Government portals are now juicy, recurring targets, and when they fall, businesses pay in confusion and late fees. The cheapest insurance is filing early and saving a screenshot of any outage you hit.
Source: Euronews: Belgium's finance and tax agency hit by cyberattack
Anthropic's unreleased AI found and chained real flaws in Chrome, Safari, Windows, and Linux
Anthropic's Frontier Red Team published research showing that an unreleased, internal Claude model found serious security flaws in software almost everyone runs, then chained them into working attacks with what Anthropic calls little human involvement. It detailed five chains: a one-click Chrome attack that escaped the browser's sandbox on a Mac, a one-click Safari attack that could read data from other signed-in sites, a lock-screen bypass that pulled photos and personal data off a locked Pixel 9, a remote crash bug in Windows servers, and a no-login takeover of Linux servers that took the model 48 minutes to set up and five more to reach a working command shell.
Anthropic stresses that these were lab demonstrations, not real-world attacks, and that every flaw was reported to the vendor first under its coordinated disclosure policy. Google has credited Anthropic for fixes, and Chrome 156 alone included 62 Anthropic-reported flaws, all rated critical or high. Anthropic is not releasing the model, and is instead expanding a program that gives vetted security professionals access to its existing models for defensive work. The Hacker News noted that Anthropic does not break down how much of each chain was done by humans.
In short: Anthropic says an unreleased Claude model found and chained real, now-patched flaws in Chrome, Safari, Android, Windows, and Linux into working exploits.
What it means for your business: The flaws are fixed, but only on devices that actually install updates. AI is making it faster to find these bugs, so the gap between "patch released" and "patch exploited" will keep shrinking.
My take: This is my one technical pick today, and it is worth it. A year ago, chaining four browser bugs into a working attack took elite human teams weeks. Now a model does it in the lab with light prompting. The good guys got there first this time. The lesson for owners is simple: turn on automatic updates for browsers, phones, and computers, and stop clicking "remind me tomorrow."
Hackers plant code in hundreds of Adobe Commerce and Magento stores
Attackers have breached hundreds of online stores running Adobe Commerce and Magento, the software behind a huge number of e-commerce sites, according to security firm Sansec as reported by The Hacker News on Wednesday. They are exploiting a critical flaw nicknamed PolyShell (CVE-2026-48356, rated 9.3 out of 10) to inject a hidden loader into a store's content blocks, which can then pull down and run whatever code the attacker chooses. The activity has been going on since at least July 2025.
The US Cybersecurity and Infrastructure Security Agency (CISA) added the flaw to its list of vulnerabilities known to be exploited on Tuesday, giving federal agencies until November 10 to fix it. No affected stores have been named, and researchers have not recovered the final payload, so what the attackers ultimately want is still unclear. On shopping sites, hidden code like this is often used to steal customers' card details at checkout, but that has not been confirmed in this campaign.
In short: Attackers are exploiting a critical flaw to plant hidden code in hundreds of Adobe Commerce and Magento stores, and CISA has flagged it as actively exploited.
What it means for your business: If your online store runs on Magento or Adobe Commerce, ask your developer or host today whether you are patched and whether anyone has checked for injected code. Hidden scripts on a checkout page can cost you customers' trust and card-processing privileges.
My take: Plenty of small retailers had a Magento site built years ago and have not thought about it since. That is exactly the kind of site this campaign feeds on. Ownership does not end at launch: if nobody is patching your store, that is the first thing to fix, ideally before the holiday rush.
Source: The Hacker News: Adobe Commerce PolyShell attacks breach hundreds of stores
🧰 Tools for builders and business AI
Anthropic shows Claude Enterprise customers what their AI spending is buying
Anthropic on Wednesday launched reporting tools for Claude Enterprise customers that try to answer the question every CFO is now asking: what are we getting for this? Team reports let managers review a team's work over a chosen date range, combining usage analytics with short employee surveys to show what the team produced, what it cost, and where workflows broke down. The reports also suggest ways to improve results.
A third piece adds pooled group budgets, so a team can share one monthly token budget instead of managing spend person by person. Anthropic says it will not use this data to train its models by default. TechCrunch noted the launch comes as Anthropic heads toward a planned IPO and as companies face pressure to justify fast-rising AI bills.
In short: Anthropic added team reports, improvement recommendations, and shared budgets to Claude Enterprise so companies can tie AI spending to actual output.
What it means for your business: Whatever AI tools you pay for, you should be able to answer "what did this produce last month?" If your vendor cannot show you, start tracking it yourself.
My take: This is a useful feature and a smart commercial move at the same time, and both can be true. AI spending is the line item that grew 10x without anyone budgeting for it. Even if you are not on Claude Enterprise, borrow the idea: once a quarter, list your AI subscriptions next to what each one actually saved or produced, and cut the ones you cannot defend.
Source: TechCrunch: Anthropic just launched tools to show companies what their AI spending is buying
GitHub Copilot Fusion goes live: choose an effort level, not a model
GitHub made Copilot Fusion generally available on Wednesday across all Copilot plans. Instead of picking a specific AI model for each task, you choose an effort level (Standard for everyday coding, High for harder problems, Max for the most demanding work) and Copilot routes the work to suitable models. It works in VS Code, the GitHub Copilot app, the Copilot command line tool, and JetBrains IDEs, with more coming.
There is no separate fee for Fusion; you pay each model's standard rate, so a workflow that uses two models bills for both. For a limited time, paid plans get a 30% discount on model costs when using Fusion. Admins keep control: Fusion respects organization model policies, and individual models can be switched off.
In short: GitHub's Copilot Fusion now lets developers pick an effort level and have Copilot choose and combine models automatically.
What it means for your business: If you or your developers use Copilot, this removes one more decision from the day and may lower costs during the discount window. Just keep an eye on usage, since multi-model workflows bill for every model involved.
My take: "Pick the model" was always a weird job to hand to users, and the industry is moving toward routing it for you. That is good. The thing to watch is the bill: automatic routing is convenient, and convenience is how costs creep. Set a budget alert before you set it to Max.
Source: GitHub Changelog: Route and orchestrate models with GitHub Copilot Fusion
That is today's AI and IT news. For more context on the FEC outage and the week's other stories, see yesterday's recap.