PHP shipped urgent fixes on Wednesday for two critical flaws, nicknamed PHP2Shell, that let an attacker run code on a web server with one unauthenticated request(opens in a new tab). Every PHP installation serving web traffic is affected, and Apache with mod_php is the most exposed. News of the fix leaked on Tuesday, and the PHP team moved the release up a day, from Thursday to Wednesday. W3Techs, per Reuters, puts PHP behind nearly three-quarters of websites, and by Wednesday the Dutch cyber agency, the Port of Rotterdam, France's Interior Ministry and Finland's government network operator had taken sites offline(opens in a new tab) rather than wait.
The AI part is the news. These flaws are old. Deseret News reports(opens in a new tab) they were flagged long ago and set aside because exploiting them looked too hard. Ryan Austin, a researcher at the security firm Calif, uncovered the problems in an August audit, with help from OpenAI's GPT-5.6 Sol. Wiz, which published the technical analysis, says frontier models substantially reduced that barrier. This is the second time in three months. In July, wp2shell(opens in a new tab) gave a similar unauthenticated takeover of default WordPress installs, also found with GPT-5.6 Sol. Wiz found 25% of organizations using WordPress exposing a vulnerable server at publication, and still 10% a day later.
Your exposure is the PHP nobody listed: marketing microsites, WordPress blogs, acquired brands, agency-hosted campaign pages, vendor portals. The patch takes minutes. Finding every server takes days if no inventory exists. Ask the CISO for three numbers today: internet-facing PHP servers, how many are patched, and how many a third party runs. Then decide who may take a public site offline, as several European governments did, and what an hour offline costs, so the call is not improvised next time.
The budget point is the patch window. Windows measured in weeks were set when writing an exploit was slow and expensive. If models keep cutting that cost, the gap between disclosure and mass attack shrinks to hours, and a leak like Tuesday's removes even that. Fund the dull parts: an asset inventory that covers third-party hosting, automated patching for web runtimes, and a web firewall that can push a virtual patch the same day. The same models that found these flaws can audit your own code. Point them at it before someone else does.
Action items
Frontier models just made a decades-old flaw in the language behind most of the web exploitable, and the details leaked before the patch.
For the CISO
Count every internet-facing PHP server, including ones a vendor or agency runs for you, and patch or take offline today.
For the CIO
Agree who has authority to take a public site offline, and what an hour of downtime costs, before the next disclosure.
For the CFO
Patch windows set in weeks assume exploits are expensive to write. Fund inventory and automated patching on the assumption they are not.
For the CAIO
Point the same models at your own code. Defenders and attackers now have the same tools for finding old bugs.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- Help Net SecurityPHP2Shell: Critical PHP vulnerabilities patched, update now!(opens in a new tab)
Fixed in PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, with source backports for 7.4 to 8.1. Linux and LTS vendors had patches before release, so check their packages first. The UK NCSC urges organizations to consider taking offline anything they cannot patch now.
- RTÉWebsites go dark as AI exposes critical security flaws(opens in a new tab)
Dutch, Danish, Finnish and French public services went offline as a precaution, and Wikipedia's operator pulled systems. Taking a site down is a valid control. Agree in advance who can order it and what the downtime costs.
- Infosecurity MagazineOpenCode Breach Could Affect OpenAI, Anthropic and Google(opens in a new tab)
An attacker reached a legacy system holding model provider keys for 1.4% of OpenCode Console users. Any tool that stores your model keys is a billing and data risk. Give it a restricted key with a spending cap, never a production key.
- TradingKeyEquifax (EFX) Q3 2026 Earnings Call: AI Margin Gains Offset Mortgage Weakness(opens in a new tab)
Equifax doubled its multiyear AI productivity target to $150 million from $75 million set in February, and raised 2026 margin guidance, crediting cost control and AI together. A named dollar target gives your CFO a public benchmark.