Developing · updated 2026-10-22T08:20Z
The PHP flaws: AI found them, a leak forced them out, and the web went offline
Early on 22 October 2026, governments told organisations to take websites offline if they could not patch two critical flaws in PHP. PHP is the language behind WordPress, Wikipedia and a large share of the web. The flaws were found by researchers using frontier AI, and a fix was due in late October. Then the details leaked and the disclosure was rushed. This section follows the story as it develops, and it keeps the evidence: every source captured as we found it, dated and hashed, so that what was said today can still be checked when the pages change.
Developing, and agent-produced. This page is assembled by software agents working from captured sources, and the editor is Dinis Cruz. Most of what is below is reported, often second-hand, and it is marked that way. Nothing here is advice on what to patch. For that, read the PHP project and your national cyber-security agency.
What we know so far
- The flaws are in the PHP language itself. The NSA's cyber division calls them
two critical vulnerabilities
, and they could let an attacker run their own code on a web server without authenticating. BBC News - The UK's NCSC, the NSA and their counterparts in the EU and Australia advise patching.
Where immediate remediation is not feasible
, they advise taking the system offline. BBC News - Researchers at Calif found the flaws in August with OpenAI's models, under OpenAI's Patch the Planet initiative, which the PHP Foundation joined in July. Wiz and OpenAI found them independently shortly after. BBC News
- The coordinated disclosure broke down when basic details leaked. According to the PHP Foundation, as the BBC quotes it, Calif then
mistakenly published a video
. That forced the announcement forward. BBC News - Estimates of the scale differ:
something like 30 million websites
(Alan Woodward) againstroughly 70%
of the internet (Wiz). The BBC says it could not verify the second figure. BBC News - Sites reported as affected or taken offline include parts of Wikipedia, Edinburgh council, French government services, the University of Geneva and the BBC. BBC News
- As of 08:00 BST, the researcher who found the flaws said there was no evidence of exploitation yet. Probing has been reported (Amazon, Cloudflare), but only second-hand so far. BBC News
What we do not know yet
- The CVE numbers, the affected PHP versions and SAPIs, and the fixed releases. The primary advisory is being captured now.
- Who leaked the details, where and when, and what exactly the deleted video showed. So far the accounts come from one party.
- What the 70% and the 30 million figures are measured from, and why they differ so much.
- Whether any site was actually exploited before it was taken offline or patched.
- How the AI-assisted discovery worked. The researchers call it
a complex chain of steps, at the edge of what human researchers could practically achieve
. What did the model do, and what did the people do? - What this says about coordinated disclosure once AI can find flaws, and turn a leak into an exploit, faster than a patch can be shipped.
The timeline
Each event is dated as its source states it. Where we worked a date out ourselves, it is marked inferred.
| When | What | Source |
|---|---|---|
| 2024 | PHP Foundation researchers find a theoretical flaw in PHP but cannot work out how an attacker could use it | BBC News |
| 2026-07 | PHP Foundation accepted into OpenAI's Patch the Planet initiative (frontier AI to find and fix flaws in popular software) | BBC News |
| 2026-08 | Calif, part of the project, finds a serious flaw in PHP; researchers at OpenAI and Wiz find it shortly after | BBC News |
| 2026-08/2026-10 | PHP Foundation works with the companies on a patch, planned for release in late October | BBC News |
| 2026-10-20 (inferred) | Coordinated disclosure breaks down ('worked well until two days ago', per the PHP Foundation as quoted by the BBC): basic details leak on social media; Calif 'mistakenly published a video' date inferred from 'two days ago' in a 22 Oct post | BBC News |
| 2026-10-22T06:15:40Z | BBC opens its live page: governments and companies rush to protect websites | BBC News |
| 2026-10-22T06:55:50Z | BBC reports the NSA's cyber division post on X: prepare 'for an update addressing two critical vulnerabilities'; take systems offline if an immediate update is not feasible time is the BBC post, not the NSA post | BBC News |
| 2026-10-22T07:00:00Z | Ryan Austin (Calif) on Radio 4 Today: no evidence of exploitation yet ('Not at the moment') 08:00 BST post, seen in capture 20261022T080453Z | BBC News |
The claims, and who makes them
A claim stays reported until a primary source confirms it, and unverified where even the reporting outlet says it could not check it.
| Claim | By | Status | Source |
|---|---|---|---|
| Wikipedia (parts), Edinburgh council, French government sites, the University of Geneva, online shops and the BBC affected or taken offline | BBC | reported | BBC News |
| Found using frontier AI (ChatGPT) by Calif, then Wiz and OpenAI; 'a complex chain of steps, at the edge of what human researchers could practically achieve' | Calif, Wiz, OpenAI | reported | BBC News |
| Attackers could take control of affected web servers without authentication and steal data | Paul Spain, Gorilla Cyber Security; Woodward | reported | BBC News |
| Calif 'mistakenly published a video' showing basic details of how the flaws could be used Calif's own account needed | PHP Foundation (as reported by the BBC) | reported | BBC News |
| The 'most severe' flaws ever detected by AI | Wiz | reported | BBC News |
| No evidence yet that hackers have exploited the flaw | Ryan Austin, Calif (08:00 BST) | reported | BBC News |
| Where immediate remediation is not feasible, take the system offline until fully mitigated | UK NCSC; NSA; EU and Australian counterparts | reported | BBC News |
| Amazon and Cloudflare saw sites being probed for the flaw second-hand; need AWS or Cloudflare statements | Alan Woodward | reported | BBC News |
| Something like 30 million websites could be affected | Alan Woodward, Surrey Centre for Cyber Security | reported | BBC News |
| Roughly 70% of the internet is affected BBC: 'unable to independently verify that figure' | Wiz | unverified | BBC News |
| Two critical vulnerabilities need the NSA post itself and the CVE IDs | NSA Cybersecurity (X) | reported | BBC News |
| Criminals are using AI to 'vibe code' an attack method | unnamed researchers (BBC) | unverified | BBC News |
How the evidence is kept
Every source is kept in a private, encrypted sgit vault that several agents and the editor write to. For each source the vault holds dated captures: the page as served, the page after it ran in a browser, its text, a full-page screenshot and the HTTP headers. Each file's sha256 is recorded at the moment of capture. Captures are only ever added, never edited or deleted, so a live page that changes leaves a series. Each source also has a short note on what it says and who says it, and every claim on this page points back to the sources that make it.
The captures are third-party pages, so they are not republished here. This page links to the originals and quotes them briefly, and the sources page lists every capture with its hashes. Anyone the vault is later opened to can check that the bytes are the ones we saw, on the day we saw them.
Every source, every capture and its hashes →
For an agent
The data behind this page is data/sources.json, data/timeline.json, data/claims.json and the editor's data/story.json. php-flaws/build/build.py --from <vault clone> re-exports the vault's metadata and rebuilds both pages. The captures themselves are in a private sgit vault and are not published; each one's sha256 is listed on the sources page.