Something happened in the second half of October 2026 that most people have not taken in yet. It was not one firm having a bad week, but a public show; in date order, of how cheap it has become to break into a computer you own.

So let me walk through what happened, because the dates matter. Then I want to talk about what those articles leave out. I mean your own kit: the laptop, the desktop, the router, and the small server you put on the internet.

What happened, in date order

On 14 October, banking watchdogs in South Korea said that 16 finance firms had been hit by attacks they suspect were run by AI agents. The count was up from 7 reported earlier, and about 100,000 records were exposed, such as national ID numbers.

On 19 October, Reuters reported on one Chinese-speaking operator who had broken into sites owned by more than 100 firms, among them Deere, SharkNinja and Cummins. He did very little of it himself. His direction to the agents ran to under 100 Chinese characters, and they probed and broke in over hours. The whole campaign used about US$5,000 of model tokens, which works out at about US$50 per successful break-in.

On 21 October, PHP shipped emergency fixes for 2 holes, tracked as CVE-2026-109618 and CVE-2026-109619. PHP runs about 70% of the world’s websites. The Wiz team that found the bugs said they “rank among the most severe vulnerabilities discovered in PHP, and arguably in software history”. Government agencies told firms to take sites offline until they could be patched.

Also on 21 October, Microsoft published its account of Storm-0940, a China-state aligned group that has used an AI attack framework called ARTEX since at least September. Microsoft says it went after hundreds of firms, and reached full access to a cloud account in under 3 hours in several cases. It also ties the group to the ConnectWise Cloud Backup breach, which matters later.

On the same day, Railway, a cloud platform used by a great many developers, said an outsider may have read its customers’ scrambled settings, and unscrambled some of them. Those settings are where API keys and database passwords live.

On 22 October, CrowdStrike put out a report on 4 groups running agentic campaigns. Those are attacks where the AI makes choices and runs commands, rather than just writing code for a human. Its best sentence is the one to remember: the main limit on scaling this is “not engineering complexity but token budgets”.

Why this week is different

I spent years inside corporate cybersecurity, and I remembered the “script kiddies” in the mid-1990s. Most of us filed them under noise: someone else’s tool, run badly, usually caught. The mistake was never our estimate of their skill. It was the belief that the number of people able to attack anything would stay small.

We have had hacking scares for 30 years, but 3 things changed at once this month. First, finding the hole became cheap. Anthropic reported on a project it runs with about 50 partners, where its models found more than 10,000 serious bugs in important software. Cloudflare found 2,000 bugs in its own core systems, and 400 of them were serious. Those bugs were always there, and nobody had the hours to go looking.

Second, using the hole became cheap, and the Reuters story is the proof. One person, about US$50 per break-in, broke into more than 100 firms. CrowdStrike found that even groups with crude, off-the-shelf setups broke into many firms, because the limit is now the AI model bill rather than the skill.

Third, fixing has not sped up at all. Anthropic says the bottleneck is the human capacity to sort, report and patch, which is the most worrying line in its own research. An attacker has to be right once, while a patch has to be written, tested, installed and checked, and then survive the person who is already doing 3 jobs.

The part you CAN actually fix

Almost all the advice this week was to update your phone and stop clicking links. That is fine, BUT it is not enough. The danger now is that automated tools scan the whole internet for anything that answers, and a great deal of what answers belongs to people like us.

In August 2026, 5 government agencies in the United States put out a joint warning about industrial controllers. The detail worth reading twice is how the attackers find them. They use internet scanning services to find exposed devices, and they dress their attack scripts up as monitoring tools. Swap “controller” for “NAS”, “camera”, “router” or “the dashboard you opened so you could check on the house from work”, and the line still holds.

So the question is not whether you trust yourself to patch. It is whether you know what you are putting on the internet right now. Most home setups I have seen, mine included at one point, have at least 1 of these:

• A port forward for something you set up in an evening and meant to secure later.
• An admin page for a NAS, hypervisor or camera that answers on the public internet.
• Plug and play left on, quietly opening doors nobody asked for.
• A smart device with a default password, or firmware from 2021.
• A self-hosted app that was patched once, when it was set up.
• A backup or remote-access tool. This is the ConnectWise lesson: the thing you put in to protect yourself holds passwords and reaches all of it, so it is a target too.

None of that needs carelessness, because it needs only having been busy, which is the normal state of a human being.

What this looks like from a small homelab

I run a monitor on my own network, and I have written about what happened when I read its output. It reported 6,688 findings, and 12 of them were real. That ratio is not a failure of the tool, because it is what scanning looks like when the scanner has no idea what you actually use.

The lesson was not to buy a better scanner. It was that the count of findings is noise, while the short list of things reachable from outside is the signal, and only that second question is urgent this week.

The other habit that earned its place here is dull and cheap. Anything I put on the internet lives in its own machine, with its own firewall rules, and it cannot reach the machines that hold what I could not replace. When one thing gets broken into, I want the damage held inside a box I can rebuild from a snapshot.

What to do this week, in this order

1. Count what you expose. Read the port forward list on your router, then the plug and play page, then the routes your cloud tunnel publishes. Write the list down, because you cannot protect what you have not listed.
2. Close what you do not need, today. Anything you have not touched in a month is a door standing open for no reason.
3. Keep admin pages off the public internet. If you need remote access, use a tunnel or a VPN that checks who you are, rather than a forwarded port with a login box on it.
4. Patch the dull layer first: router firmware, the NAS, the hypervisor, the browser, the phone and the desktop. Then the app containers on top. Reboot when it asks, because this is the step people skip and it is the step that counts.
5. Switch on two-factor login for email before anything else, because email is the reset path for every other account you own. Then do the same for the bank.
6. Check your backups, and check that one copy cannot be reached from the machine it is backing up. Then restore something small from it, so you know it works rather than believing it does.

One thing not to do: do not empty your bank account over this. Panic is a business model, and the scammers who always follow a week of headlines are counting on you to lose your head.

What I cannot tell you

I cannot tell you that patching is enough, because 2 of this week’s stories were about holes nobody had fixed yet. I cannot promise your internet provider or your router maker will warn you, because they usually do not. And I cannot tell you a cloud service you are not allowed to inspect is safe.

What I can tell you is that the gap between being a victim and being a story is usually 3 choices, all made long before the attack. What did you expose? What did you patch? What was really in your backup?

That is worth an hour this weekend, not because the internet is ending, but because this week shows what the new normal costs. The bill lands on whoever left the door open.

Sources and corrections: this piece is built from primary sources, listed with links in the claim sheet beside the draft. One widely shared claim from the original X posting essay that prompted it, that a state-aligned group went after more than 2,000 firms, is not what Microsoft’s own write-up says. It says hundreds, so I used hundreds. Two claims I could not read at source were left out rather than repeated. But nevertheless, it warrants the highest importance for me to want to share this with everyone, right now.