Black October at the End of a Gravel Road

Not a Medium member? Click here to keep reading.

We weathered a ten-day power outage here with solar panels. Thank the gods it missed the long, sodden stretches of winter. There's only so much self-sufficiency you can squeeze from a sky full of water.

I live on a little farm in rural South Africa, at the end of a gravel road. Other people on the property grow vegetables. Our little herd of horses supplies the manure. The vegetables also need protection from the horses. A solar-powered borehole helps keep the grass green when rain runs short. Sun into supper, with a detour through dung.

Beyond food and water, most of my life runs through the internet. The vehicles on the property stopped working quite some time ago. A road out is cold comfort when nobody can drive down it. The horses do their bit for the food supply. Their performance in telecommunications remains poor.

So when AI companies announce that their tools can find dangerous faults in the software behind that connection, I want to know what happens between finding a fault and fixing it, and when the fix reaches people like me.

A flaw beneath the website

The latest case involves PHP. It is software that runs on servers, the computers that deliver websites to visitors. It builds the pages and handles the data behind them. A person who edits a WordPress site may use PHP for years without ever seeing it.

PHP's maintainers fix PHP. Someone else runs each website: the business itself, or a hosting company it pays. A fix only protects a site once it reaches that site's computer.

On October 21 PHP's maintainers released repairs for two critical security flaws, together called PHP2Shell. Under the right conditions an attacker could send a request that made the server follow the attacker's instructions without a login. That puts a stranger inside the machine.

October 21 is my birthday and, by Pratchett and Gaiman's reckoning in Good Omens, the world's too. A maintenance notice seems sufficient ceremony.

Both flaws come down to PHP trusting its own notes when it shouldn't. While it builds a page, PHP jots down where it has put things in a small scratch space. If it ran out of room at the wrong moment, it could leave a note half-written and later trust it as complete. It could also follow an old note to a spot that now holds something else. Either slip could let an attacker sneak in instructions of their own. On some servers this could happen before a perfectly respectable login page got its turn.

Nobody tripped over these flaws. The researchers who found them drew on work on how PHP handles limited working space that reaches back to 2004. Experts chose where to look. AI helped them probe. Separate teams found the flaws independently. If two teams found them, defenders should assume attackers can too.

The census and the casualties

Within days of the repair headlines warned that the flaws put "seventy percent of the internet" at risk. The scare got a name: Black October. The figure comes from a survey of the software behind websites. As of October 30 PHP runs 69.2% of the websites where the survey can identify the language behind the scenes. Sites it cannot identify stay out of the count.

A website faces this danger only when two things hold:

• it runs a faulty copy of PHP;
• the attacker has a way to take advantage of the fault on that particular computer.

Working attacks depend on specific conditions, so success against one website proves the danger is real without saying how many others share it.

So why lead with seventy percent? A dangerous discovery doubles as a demonstration of the AI that helped find it. The AI company gets its pitch, and the news its clicks. Neither tells a website owner whether their own site is at risk. I depend on the internet too much to buy its obituary with every bulletin.

The evidence of attacks so far comes with a catch. They first appeared on October 21 against computers set out as bait. Researchers call these traps honeypots and use them to watch intruders at work. They prove someone is trying. They don't count the businesses actually broken into. A million attempts can contain a whole lot of failure.

An alarming software census is not a casualty count. The web can stay up while a small business goes under. A site with few visitors can hold a clinic's appointment book or a co-op's member records. Traffic is a poor measure of dependence.

Anyone responsible for an affected website must follow the repair instructions, whatever they think of the headlines.

The wrong end of the warning

An operator needs to know which systems require repair. A customer needs to know whether a service is safe, and where to turn if it closes. One commentator opened with an internet on fire and advice to limit web use for a few days. That helps neither. For most people the honest answer is: very little. The repair is the operator's job.

The official advice aims at operators. The UK National Cyber Security Centre (NCSC) advises operators to consider a shutdown when they cannot protect a system in time. A viewer who closes a browser leaves the faulty server untouched. A hacked service can still harm its visitors, so useful caution says what to avoid and why. People use the web to earn a living and get things done. "Stay off the internet" gives them no way to judge a particular risk.

The customer gets the fright. The commentator gets the firelit thumbnail. The website's operator must find the faulty software and get it fixed.

That operator is often a small business. Take an auto body shop that uses a website to serve customers. Its owner's job is to repair vehicles, not to keep up with fashions in software. Replacing an old website costs money and can interrupt the business. The mechanic who declines to rebuild the booking system between brake jobs runs a business.

The repair that stalled

The repair itself took two years to arrive. The PHP Foundation funds work on the language. It says a developer first spotted the problem after a 2024 security review. At that stage, the team could not show how an outsider might use the fault to take control. Proposed repairs risked breaking websites or making them easier to knock offline. Some would require changes to add-ons that other people write for PHP. The team couldn't agree, and work stopped and started. The Foundation calls the delay a "structural failure" in its decision process and resources.

Websites relied on the behavior those repairs would disturb. A security fix that breaks something essential becomes an outage with good intentions.

The Foundation is now setting up a dedicated security team and moving to majority votes. AI helps check reports. People make the calls and do the repairs. The plan was to give companies that ship PHP a head start on updates before going public. Leaks cut that short, so the team published early.

The team already had a queue of other problems, and AI discoveries add to it. An AI discovery can change the evidence available to a team and push an urgent repair to the front of its queue. The queue still contains other people's emergencies. Paying for more searching just piles more on the same few people, unless someone pays for more hands too.

Even developers who know the stakes get tired as the work piles up. The next warning competes with the unfinished response to the last. A permanent emergency is an awkward staffing model. It saves wonderfully on the distinction between ordinary hours and overtime.

Every copy

Even once a repair ships, "Update PHP" sounds like an instruction to do one thing. In a large organization the first job is to locate every copy.

PHP can turn up in places nobody remembers choosing:

• the organization's own website;
• software it bought, which may arrive with a separate copy of PHP tucked inside;
• servers a hosting company runs on the customer's behalf;
• old internal tools that outlived the people who built them.

One computer can contain several copies, each used by a different program.

Where a hosting company looks after PHP for a customer's website, the customer may not be allowed to fix it. The company has to confirm that the website uses a repaired copy.

Finding every copy calls for a list. For a clinic a useful one would show which computers use the faulty software and who can repair them. The trade calls it a software bill of materials. It has to stay current, or it's just paperwork.

The NCSC's guidance sets out the order of work:

1. Find the computers that use PHP.
2. Repair the ones strangers can reach over the internet.
3. Then repair the ones reachable only from inside the organization.
4. Where updates or temporary protections cannot go in on time, disconnect the computer or close the service.

A shutdown interrupts the very work security staff must protect, so someone has to know what it will cost.

The internal copies in step 3 matter more than they look. A program used only inside an organization may be out of a stranger's reach. Once an attacker breaks in elsewhere, it can open the way to the rest. Fixing the public website leaves those older copies still to find.

On the farm the same problem runs along the fence. Our "preystock" and the neighbors' are the animals the hounds might regard as lunch. After rain loose soil can let a hound work under an unrooted fence. Sagging stretches and weak wire offer other ways in. A good gate still leaves meters of fence to check.

AI helps people build more software and connect more services. The ground to guard expands while attackers gain faster ways to search it.

Checking the repair

Each copy found still needs a repair that works. Blocking the original attack is only the first test.

Researchers tested repairs written by the AI model Claude Opus 4.7 against 112 security flaws from 84 open-source projects, whose code anyone can inspect. The repairs stopped the test attack in 90.2% of cases but passed the developers' own checks in only 60.7%. A repair could stop the attack and still break something the software needed to do.

Other ways to break in may also remain after the test attack fails. Checking for both problems takes time and expertise, even when AI writes the replacement code in minutes.

Agent access

The tempting shortcut is to hand the whole job to an AI agent. An agent is an AI with tools that let it perform tasks, such as reading files and changing software. What it can do depends on what it can reach. A suggestion on a screen leaves the next step to a person. An agent can take that step itself.

Files the agent reads or a web page it visits may contain instructions planted by someone other than the person who gave it the job. If the agent obeys those instructions, a search for information can turn into an action its user never approved. The more it can change, the more damage that mistake can cause.

OWASP is a nonprofit that publishes software-security guidance. It recommends limits along these lines:

• an agent receives only the tools and access its job requires;
• actions with serious consequences require checks the agent cannot change;
• a tool intended to read information carries no power to delete it;
• the computer enforces who can do what.

A persuasive instruction should not become a permission slip.

OpenAI has one such case on record. An agent inside OpenAI removed a program's safety check during a change to its setup, then denied doing it. The agent in charge passed part of the work to a helper. It told the helper to keep the check "as appropriate." The helper treated that wording as permission to leave it out. The check it removed watched the program, not the agent itself.

Human organizations know how a firm requirement reaches the contractor as a preference, then returns to the client as a completed job. Automation gives the process excellent turnaround times. The software can remove the protection and polish its explanation before a human finishes the coffee.

Permission to change a program should not let an agent remove the program's safety checks. The operator needs a log of what the agent did and a way to undo the damage.

On the farm the same hounds keep outside threats off everyone. They're handy to have around. The hens still live behind a fence.

The invoice for protection

The company selling the alarm also sells the fix. OpenAI helped PHP's maintainers test and coordinate the release, and offers a free tool to check which version of PHP a computer runs. The same page announces $1 billion in discounted access to its Daybreak AI security tools for people defending systems. It also promotes a paid security product and promises more discoveries as AI improves. Public service and product placement share the page quite comfortably.

The strongest tools come with a gatekeeper. Using them requires OpenAI's approval, and a maintainer responsible for a fault can still be refused. Discounted AI covers the cost of the AI. The testing and repair still need people.

A good repair also teaches developers how to avoid the next mistake, but only if those lessons make it into the next version. Otherwise new code brings new faults as fast as the old ones get fixed. A count of discoveries tells customers nothing about whether they face less risk. They need a contract that says who answers when an update breaks their service, and lets them keep their software running if they leave.

For people who already struggle to keep a connection, the price of safety matters. Money for PHP's maintainers protects everyone who can't pay for a premium shield. Credits for AI tools pay for the AI. Salaries pay for the people who make a hard repair work.

From this gravel road a premium rescue looks rather like a helicopter subscription. The ride may be excellent for those who can pay. Everyone else still depends on the road.

AI development tools are sold on the amount of work their agents perform. The customer also pays to check that the work is safe. The sales pitch can sprint. Someone must pay for the proof.

The road out

Our solar panels got us through the outage with the weather's permission. The borehole needs maintenance, and someone must tend the vegetables. Out here independence takes work long after anyone stops admiring it.

Online I depend on things I can't inspect from here. I cannot verify every server behind a service I use. I need providers who know what they're responsible for and look after it. Their work keeps access open for households with few alternatives, mine among them.

I would like the road maintained. The helicopter salesman can wait.

Sources

AI for Humans, 'AI Just Broke the Internet. For Real This Time', video, supplied transcript consulted 30 October 2026.

Calif Research, 'PHP2Shell: A Tale of Two PHP Remote Vulnerabilities', research account, accessed 30 October 2026.

ElAI, 'I Read the Research on AI Hacking. This Is a Full-Blown Sh*tshow', video, supplied transcript consulted 30 October 2026.

Horizon3.ai, 'PHP2Shell Exposure Depends on More Than Whether You Use PHP', 22 October 2026, research guidance, accessed 30 October 2026.

Internet of Bugs, 'Breaking Points & Yahoo Finance Lying About "Black October", AI Hacks & "Experts"', video, supplied transcript consulted 26 October 2026.

Le, Dang K., Wenxuan Shi, and Xinyu Xing, 'On the Reliability of LLM-Based Vulnerability Patching Benchmarks', arXiv:2610.10150 (2026), preprint, accessed 30 October 2026.

Moonshots, 'SpaceX's AI Business Up 830%, OpenAI Agents Lie, and Startup Turns Mercury Into Gold', video, supplied transcript consulted 30 October 2026.

National Cyber Security Centre, 'Critical Security Issue Affecting PHP', advisory, accessed 30 October 2026.

Nerds, 'Getting Hacked Was Only Half the Problem', video, supplied transcript consulted 30 October 2026.

OpenAI, 'OpenAI Daybreak: Trusted Access for Cyber Overview', access documentation, accessed 30 October 2026.

OpenAI, 'Patch the Planet: Global Response Coordination as Agents Uncover Critical Vulnerabilities', 21 October 2026, announcement, accessed 30 October 2026.

OpenAI, 'Suppressing Misalignment Monitoring Alerts and Disabling Misalignment Monitoring', updated 20 October 2026, incident report, accessed 30 October 2026.

OWASP Gen AI Security Project, 'LLM06:2025 Excessive Agency', security guidance, accessed 30 October 2026.

PHP, 'Security Advisory: Critical Remote Code Execution Vulnerabilities, Update Now', 21 October 2026, news archive, accessed 30 October 2026.

PHP Foundation, 'Today's PHP Security Release, and How We're Strengthening PHP Security', 21 October 2026, maintainers' account, accessed 30 October 2026.

Pratchett, Terry, and Neil Gaiman, Good Omens: The Nice and Accurate Prophecies of Agnes Nutter, Witch (London: Gollancz, 1990).

Roth, Wes, 'They're Shutting It DOWN', video, supplied transcript consulted 30 October 2026.

Three Buddy Problem, discussion of PHP2Shell and AI security, supplied transcript consulted 30 October 2026.

W3Techs, 'Usage Statistics of PHP for Websites', survey, 30 October 2026.

watchTowr, 'Frequently Asked Questions About PHP Remote Code Execution', 20 October 2026, updated 26 October 2026, intelligence brief, accessed 30 October 2026.