Black October: AI turned an old PHP flaw into an emergency, and is helping to fix it

On Wednesday 21 October, the people who maintain PHP released an emergency fix and told everyone using it to install it immediately. Within hours, the Finnish government, Edinburgh council and the organisation behind Wikipedia had switched off parts of their websites rather than leave them online.

Some security researchers have started calling it "Black October". Much of the coverage has told it as a story in which AI rescued the internet. The BBC's explainer, for example, says the flaws had gone undetected for more than a decade and that "it took AI to bring them to light."

The real story is more tangled than that. The weakness had already been spotted, in part, two years earlier; what AI did was show how to turn it into a working attack. The same kind of tools are now expected to help attackers move faster, and they are also helping defenders check the fix and spot attacks. AI is both a cause of this emergency and part of the cure.

That matters for any firm whose website or client systems run on PHP, because the right response depends on understanding which part of the story applies to you. This piece sets out what happened, what AI actually did, what is proven and what is not, and what a small firm should do.

What has happened?

On 21 October the PHP project released new versions, numbered 8.5.12, 8.4.27, 8.3.36 and 8.2.35. According to the PHP project's announcement, they fix "two critical vulnerabilities that can allow unauthenticated remote code execution". In plain words, an outsider who has never logged in could make a website's server run commands of their choosing. The project added that "there are no complete mitigations, updating is the only safe choice."

The release came a day earlier than planned. The PHP Foundation, the non-profit that maintains the language, says in its account of the release that details began to leak before the agreed date, including the flaws' official reference numbers on 20 October, so it brought the release forward.

The UK's National Cyber Security Centre (NCSC), the government body that advises organisations on cyber security, issued an alert the same day. It says that "organisations using any version of PHP are affected". It tells organisations to find every system running PHP and update it. If they cannot update, or apply temporary protections quickly, it says to consider taking the system offline. It also tells them to check for signs that someone has already got in and left a hidden way back in, known as a backdoor.

Many organisations followed that advice. The BBC reported that the Finnish government, Edinburgh council, parts of the BBC and the Netherlands' cyber security agency took websites offline, and that Wikimedia, the organisation behind Wikipedia, switched off some systems as a precaution. On 22 October the UK technology minister, Baroness Lloyd of Effra, told the House of Lords that a number of government services were unavailable, and that the government was "not currently aware of any confirmed compromise of central government systems", according to The Register.

What is PHP, and why does a flaw in it matter so much?

PHP is a programming language that runs on the computer that hosts a website, called the web server. When someone visits a page, PHP collects the right information and builds the page they see. W3Techs, which surveys websites, finds PHP behind 69.3% of the sites whose server language it can detect. The BBC lists Wikipedia, online shops and WordPress, the system behind many websites, among the things that use it. Many client portals, booking systems and online forms are built in PHP too.

Most website flaws sit in one product, such as a single WordPress plugin. These sit in PHP itself, so they reach every site built on it, whatever that site's own code looks like. Researchers at the security firm Wiz, one of the teams that found them, named the flaws "php2shell". In their write-up they showed that one of them can leak passwords, login tokens, the keys that connect a site to other services, and security certificates from the server's memory.

Because the flaws are in PHP rather than in WordPress, updating WordPress, its theme or its plugins does not fix them. Cloudflare, which protects a large number of websites, puts it plainly: "updating website code or a plugin(s) alone does not fix these vulnerabilities." The fix has to be applied to PHP, which on most small firms' sites is managed by the hosting company or the web agency.

Did AI save the day?

Not exactly. AI played three different parts in this story, and only one of them is the rescue the headlines describe.

First, AI made a known weakness usable

The weakness was not completely unknown. The PHP Foundation says one of its developers saw the first signs of the underlying problem while fixing issues raised by a 2024 outside security audit. At the time nobody could show how it might be used in an attack, so it was treated as a theoretical concern, and the fix stalled. The Foundation calls the delay "a structural failure, caused by both the security team's decision process and its lack of resources."

What AI changed was the step from theory to a working attack. According to the Foundation, a researcher at the security firm Calif reported a way to take over servers on 11 August, OpenAI independently reported the same root cause, and Wiz followed on 24 August. Wiz says it used AI models from Google and OpenAI to find the flaws and to build working attacks against real software such as WordPress. In its own words, "Frontier AI models substantially reduced that barrier, turning a longstanding flaw into an exploit." It adds that "a single researcher completed in a week what previously would have taken a world-class team months of work."

So AI did not just reveal a hidden problem. It turned a known, hard-to-use weakness into something that could be used against real websites. The researchers who did this reported it privately, which is why a fix existed by 21 October. But it is also why the flaw became an emergency.

Second, AI is expected to speed up attackers

Once the details of a flaw are public, the same tools that helped researchers can help criminals. Hong Kong's computer emergency response team, HKCERT, warned in its advisory that "attackers may also develop attack tools more quickly", because AI is lowering the technical barriers to building them.

That is a warning, not yet a finding. We have not found a report showing that attackers used AI against php2shell itself. What is shown is that attackers were trying the flaw early. Amazon told The Register that its decoy servers, set up to attract attackers, recorded an attempt on 17 October, four days before the fix came out.

The same week also showed AI being used in a separate attack. On 21 October Microsoft described a state-aligned group it calls Storm-0940, which used ARTEX, an open-source testing framework in which AI agents carry out an attack step by step on their own. The group used it to scan for weaknesses and sign in to organisations' systems at machine speed. Its main route in was ConnectWise, a company whose cloud tools many IT support providers use to manage their clients' systems: Microsoft assesses that the group compromised some ConnectWise cloud applications and used their identities to get into customers' Microsoft 365 accounts. Microsoft says that "the majority of the targeting wasn't successful", but that in some cases the group gained high-level access in under three hours.

Third, AI is helping defenders

AI is also part of the response. Wiz says Google's Gemini model contributed not only to finding the flaws but to detecting attacks and checking that the fix worked. The reporting itself was a defensive act: because the researchers told the PHP team first, the fix was written and tested before the details were widely known.

Put together, AI made the old PHP flaw usable, then helped fix it. The same capability sits on both sides. That is why the "AI saved us" version of the story is misleading: the next time, the people using AI to find a flaw may not report it.

Is it as bad as it sounds?

The flaws are real and rated critical, the fixes are out, governments have told organisations to act, and attackers were trying the flaw before the fix existed. Amazon advises organisations to look for suspicious activity back to at least 17 October.

What has not been shown is that large numbers of websites have actually been broken into. Most of the sites that went offline did so as a precaution. The security firm Arctic Wolf, which saw widespread scanning after the disclosure, reports "comparatively little evidence of reliable exploitation at scale." Dan Guido, chief executive of the security firm Trail of Bits, wrote in a post quoted by 404 Media that "AI accelerates exploit development. Knowing how to turn a bug into a shell still takes expertise." A shell here means control of the server. His point is that AI has made attacks quicker to build, but has not yet made every attacker an expert.

Our reading is that this week is more likely a turning point than the start of the internet breaking down. It is the week the double role of AI in cyber security became visible to people outside the industry. The change that matters for a firm is timing: in this case, attackers were trying the flaw before most website owners could have installed a fix.