AI now helps find software flaws and break into networks. Why the AGI Zeitgeist did not move

On the afternoon of October 21st, 2026, organisations began switching off their own websites. In Belgium, the Flemish government, the Flemish Parliament, the foreign ministry and the country’s own cybersecurity agency took their sites down as a precaution, according to the Brussels Times. In Canada, a minister said some Government of Canada websites were “temporarily taken offline as a precaution” (CP24). None of these reports describes a break-in. The owners were waiting for a fix.

The fix was for two flaws in PHP, the programming language that runs a large share of the world’s websites. The flaws could let a stranger take over a web server without logging in. PHP’s own advisory of that day told everyone to update now, and the UK’s National Cyber Security Centre told organisations that could not apply the stopgap measures in time to “consider taking the system offline”. Within a day, people on social media had a name for the whole week: Black October.

What drew my attention was not the websites themselves but the part AI played in finding the flaws, because I watch for one particular kind of change in AI. I keep a public tracker, the AGI Zeitgeist Tracker, which asks whether we are entering the singularity, the point where machines get smarter so quickly that the world changes faster than anyone can keep up with. A week in which AI was part of the story of a flaw in much of the web seemed like the kind of week that should move it. But despite the news, the tracker’s answer stayed exactly where it was, and I think the reasons are worth setting out, because they say something about what this week’s events really were.

What AI did this week

Two separate things happened, and they are easy to run together.

The first is discovery. The PHP flaws were found in August by two security firms working separately. Calif says it found the first way in on August 1st, 2026 and reported it to PHP’s maintainers on August 11th. Its researcher, Ryan Austin, writes that he “asked my LLM companion for help”, meaning an AI system that works with text, and that nothing special was used: no custom setup and no model that was not “generally available to everyone”. He also writes that he chose a part of the code he already knew well. Wiz says it found the first of the flaws on August 19th, 2026, using the newest and most capable AI models and its own tools, and that “a single researcher completed in a week what previously would have taken a world-class team months of work.” By Wiz’s account, every version of PHP released in the last decade has at least one of the two flaws.

The second is attack. On October 21st, 2026, Microsoft published its report on a group it calls Storm-0940, which it describes as “China-state aligned”. Microsoft’s report says it “assesses” that the group was behind the break-in at ConnectWise, a company whose software other companies use to manage their computers. The group used ARTEX, which Microsoft’s report calls “an open-source penetration testing framework”: a kit, published for anyone to download, in which several AI agents work together to break into a network. With it, according to Microsoft’s report, the group tried to sign in to more than 2,250 organisations’ accounts, and in several cases went from first contact to privileged access, meaning control well beyond an ordinary user’s, in “less than 3 hours”. Wiz added its own report on ARTEX the next day, saying what it had seen “overlap[s] with Microsoft’s recent report”, and that the kit’s developer stopped publishing its code on October 8th, 2026.

For most of the other incidents grouped under the name “Black October”, nobody has shown that AI played any part. SecurityToday went through the list and found that many of the attacks had been running for months, and that “the week’s warnings do not attribute the surge itself to AI.” Even the name is shaky. Yahoo’s report says it was coined in a post on X early on October 22nd, 2026, and that its newsroom searched twice and could not find that post.

Why the tracker did not move

My tracker answers its question with a test in four parts, and all four have to cross their lines before it will say we have entered the singularity. The first is how fast AI’s abilities are growing, measured by how long a task AI can finish on its own. The second is what forecasters expect. The third is whether AI now does most of the work of improving AI. The fourth is whether the world’s economy has changed gear. On October 23rd, 2026, the tracker’s front page puts the average of the parts that can be measured at 32 percent of the way, and its answer is “Not yet.”

None of this week’s events touches those four measures. An AI that helps a person find a flaw in PHP is doing skilled work, but it is not improving AI, and the work does not show up on the measure of task length. A state-aligned group using AI agents against companies tells us something important about how the technology is spreading, but it is not a measure of the economy.

The tracker also keeps a page of dated predictions made by the people who run the AI companies, and two of them are about hacking. Both were made by Dario Amodei, who runs Anthropic. In June 2026 he said that within twelve months, AI with the hacking abilities of his company’s most capable unreleased model “may just be available for, for anyone to, to download”. In September he wrote that he worried a swarm of AI agents could, within six to twelve months, be “capable of taking over the entire internet with a persistent botnet”, meaning a network of hijacked computers kept under one controller. ARTEX was free to download until its developer closed its code, but nobody has shown that the models it runs on hack as well as the model Amodei meant. And trying 2,250 organisations’ accounts is a long way from taking over the internet. So neither prediction has come true. Going by his own words, the first falls due around June 2027 and the second between March and September 2027, and the tracker will judge both at the end of 2027.

That does not make this week small. The tracker was built to ask a narrower question than “is AI doing anything alarming?”, and it should not change its answer just because the news was loud.

The window, as an argument

What the week does suggest is something the tracker does not measure, and I want to set it out as an argument, with its limits.

Finding a flaw has become cheaper. Wiz’s report says one person did in a week what a team used to need months for, and Calif’s report says it used nothing that the rest of us cannot use. Fixing a flaw has not become cheaper in the same way, because a fix has to be written, tested and then installed by every owner of every affected system, one by one. Some systems have no owner who will ever do that. PHP’s own advisory says that “PHP 8.1 and older no longer receive updates from the PHP project”, and its list of retired versions shows that support for 8.1 ended on December 31st, 2025. The PHP team did publish fixes that people can build themselves for three of the older versions, but a system still on a retired version is often one that nobody is maintaining.

If finding flaws keeps getting cheaper while fixing old software does not, there is a window in which the people who benefit most from cheap discovery are attackers going after the oldest systems. New code can be checked by the same tools before it ships. Old code that nobody maintains stays as it is, and every flaw found in it stays open.

That argument has limits, and this week shows them as plainly as it shows the danger. In the PHP case, the AI-assisted discovery went to defenders first. Both firms reported privately, and the fix came out about ten weeks after the first report. Calif’s report also says that “everyone and every lab had already done extensive AI assisted audits” of PHP, so defenders had the same tools and had been using them. The Belgian and Canadian sites went offline as a precaution, which is defenders acting quickly. I also cannot tell you how many websites still run versions of PHP that get no fix, because I have not found a count I could open and check. Even the jump in flaws being registered is hard to read. CRACI’s count of newly registered flaws runs from 7,815 in June 2026 to 14,930 in September, but the same page says the rise mixes AI-assisted discovery with more organisations being allowed to register flaws and with old records being filled in, and that “the published record cannot separate them.”

So the window is a hypothesis. It could be tested by watching whether break-ins through old, unsupported software rise faster than break-ins through software that is still maintained, and the registered counts cannot answer that.

Where this leaves the record

I argued in Insufficient Time for a Meaningful Answer that AI is not breaking free of our institutions but spreading into them. This week fits that picture. AI turned up inside security firms finding flaws and inside a state-aligned group breaking into companies, and what it found reached the people who had to decide whether to take their websites down. None of that is the singularity the AGI Zeitgeist is looking for. It is what the spreading looks like from inside a server room.

The tracker will keep its answer until one of its four measures moves. If a model that anyone can download is shown to hack as well as the best unreleased ones, the claims page will record it beside Amodei’s words. If break-ins through unmaintained software start to pull away from the rest, that is a measure worth adding to the record, and I would rather add it on purpose than let one loud week stand in for it.

Reading List & Conceptual Lineage

From Sentient Horizons

Insufficient Time for a Meaningful Answer
The argument that AI spreads into the institutions we already have, rather than escaping them. This week is an example from computer security.

Anthropic’s AI writes 80 percent of its code and leads 26 percent of its research
The previous Frontier Dispatch, on why the tracker scores one figure and not another.

The tracker

The AGI Zeitgeist Tracker. The four-part test and where each part stands.

The claims page. The dated predictions by the people who run the AI companies, including Dario Amodei’s two about hacking.

External Works

Microsoft — AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise (October 21st, 2026). The attack, and what the AI kit did in it.

Wiz — PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet (October 21st, 2026) and ARTEX in the Wild: Storm-0940 Expands Targeting of IT and Software Vendors (October 22nd, 2026).

Calif — php2shell. How the first flaw was found, in the researcher’s own words.

The PHP project — the security advisory (October 21st, 2026) and the list of retired versions.

The UK’s National Cyber Security Centre — Critical security issue affecting PHP (October 21st, 2026).

CRACI — CVE Publication Data. The monthly count of newly registered flaws, and why it cannot be read as a count of AI discoveries.