NSA Flags Two Critical PHP Flaws as Patches Ship the Same Day
The U.S. National Security Agency warned on Wednesday, 21 October 2026, that two critical vulnerabilities had been found in PHP, the server-side scripting language behind a large share of the web. The PHP Foundation issued patches for both the same day. The NSA's guidance is blunt: websites that cannot be patched immediately should be taken offline until they can be.
Both flaws are rated 9.2 under CVSS v4.0, the Common Vulnerability Scoring System, where 10 is the maximum. The PHP project announced the critical security release on 20 October, according to CERT-EU, and scheduled it for 14:00 UTC on 21 October. The project's advisory names PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 as the releases that fix the two issues. TechCrunch reports that the flaws affect every version of PHP since version 5, released in 2004. PHP 8.1 and older are no longer supported, and the recommended current version is 8.5.12.
The exposure is broad. W3Techs estimates that PHP powers roughly 70% of websites, a figure TechCrunch cites. Remote code execution, which lets an attacker run their own commands on a server they do not control, in a language with that footprint reaches well beyond the WordPress installations most readers will picture first. WordPress has stated that the flaw is not WordPress-specific, and that site operators must patch their servers and work with their hosting companies to be protected.
The discovery story is the unusual part. Calif, a security testing company taking part in OpenAI's Patch the Planet work, used GPT-5.6 Sol to find one of the flaws, and OpenAI disclosed the finding in a blog post. A Yahoo News report says Calif found the flaws in August 2026 during an OpenAI-funded audit of PHP's code. Google's Wiz team used its Wiz Atlas AI security agent, Gemini 4 Argon and GPT-5.6 Sol to find multiple PHP vulnerabilities. Wiz Research describes PHP2Shell, the headline finding of its blog post, as the first PHP vulnerability it discovered with that stack. A dated entry on the post, 20 August 2026, describes the flaw as unauthenticated remote code execution, meaning no login is required.
Patch the Planet is a Daybreak initiative that uses AI to help open-source maintainers find, validate and fix vulnerabilities, according to OpenAI's program page. Its separate Codex Security product, which OpenAI describes as an AI application security agent that analyzes project context to detect, validate and patch vulnerabilities, is in research preview, an early-access stage.
OpenAI and Wiz both say they reported the PHP flaws to the PHP project and worked on patches before disclosing them publicly. OpenAI states that researchers found the flaws before hackers did, and that it has no reports of exploitation in the wild. Its blog post frames the period between discovery and public disclosure as a "Defender's Window" for maintainers and security teams. Wiz thanked Automattic, Akamai, AWS, Cloudflare, Fastly, Imperva, Microsoft and Vercel for helping test and refine the findings.
Operational impact is already visible. The Hacker News reports that Cloudflare saw early exploit attempts, with no confirmed breaches. The BBC reported that websites were taken down worldwide as a result of the disclosure.
Wiz has warned that large language models can help adapt an exploit to different PHP versions, configurations and memory layouts. In Wiz's framing, that removes the difficulty barrier that previously made exploiting such flaws across a mixed install base expensive, since servers run many versions and setups. That is the technical claim administrators should weigh most carefully, because it describes the attacker side of the same capability that found the bugs.
In my view, the operational lesson is the timeline rather than the headline score. A 9.2 rating on a language that runs on most of the web is alarming on its face, but the practical question for an operator is how long a server has been running an unpatched release, and whether it can be upgraded from a branch that is no longer supported. Teams still on PHP 8.1 or older face a migration, not a point release, and that work takes longer than an emergency window allows.
The broader context here is that AI-assisted vulnerability research is now producing findings in mature, widely deployed open-source code, and the same tooling is available to people who do not disclose what they find. The Defender's Window that OpenAI describes lasts only as long as maintainers can ship fixes and operators can deploy them faster than exploit adaptation spreads. If the disclosure pipeline works, the same tools could shorten the life of old, latent bugs in infrastructure the web depends on. That outcome is not yet established by this episode, but it is the reason the coordinated-disclosure model matters.
