OpenAI-Backed Research Reveals Critical PHP Flaw as Companies Race to Patch

OpenAI has detailed its role in a coordinated response to two critical security flaws in PHP, as website operators face urgent software updates and evidence that attackers were testing the vulnerabilities before their public disclosure.

The company’s announcement said researchers at Calif identified a critical flaw using its GPT-5.6 Sol model through the Patch the Planet initiative. Wiz independently discovered and reported both vulnerabilities, while PHP’s maintainers led the fixes.

The immediate concern extends beyond the companies developing AI. European cybersecurity authorities have urged organizations to identify affected applications and prioritize internet-facing systems, including products that use PHP behind the scenes.

The vulnerabilities, identified as CVE-2026-109618 and CVE-2026-109619, affect PHP itself, the software used to run many websites and applications. The PHP security advisory says the flaws can allow remote code execution without authentication, potentially letting an attacker run malicious instructions on an exposed server without first logging in.

PHP’s maintainers identify installations serving web requests as at risk, including several common server configurations. They say the highest risk is associated with Apache’s mod_php configuration, where a demonstrated exploit can trigger before the website’s own application code runs.

Wiz’s technical research also describes the potential to read sensitive information from process memory, including passwords, authentication tokens and API keys. It says exploitation can occur even when the application running on PHP has no corresponding flaw in its own code.

Wiz cautions that adapting an attack to a particular deployment can require changes for different versions, configurations and memory layouts. Its researchers say AI reduced the specialist effort needed to develop and adjust those attacks, making complicated weaknesses more practical to exploit.

PHP released versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to address the vulnerabilities. Its maintainers strongly encourage immediate updates and advise organizations using distribution packages to check with their vendors, which may supply fixes through their own release channels.

The task can be confusing for smaller businesses whose websites are managed by outside providers. WordPress’s guidance explains that these are PHP vulnerabilities, rather than WordPress-specific defects, and that no WordPress update is required or planned to resolve them.

WordPress users can check their PHP version through the administration dashboard under Tools, Site Health, Info and Server. Hosting control panels may also show that information. The project advises site owners to check their host’s updates even if the provider is already preparing protections.

European authorities’ joint statement recommends coordinating with hosting and service providers. Where immediate remediation is not possible, it says organizations should consider isolating exposed applications from public networks until fixes can be applied.

Amazon Threat Intelligence reported an exploitation attempt dated Oct. 17 in telemetry from MadPot, its network of decoy systems used to observe malicious activity. That was four days before the vulnerabilities’ public disclosure.

Amazon said the attempt came from an address associated with a commercial anonymization network. It assessed that the activity suggested a threat actor, while acknowledging that it did not know how the actor obtained exploit information before disclosure.

Additional attempts appeared before the embargo lifted, and Amazon subsequently reported broad exploitation attempts in the wild. It also said it remained unclear whether the earlier attempts would have succeeded against real-world systems. Attempted exploitation should therefore not be presented as a confirmed breach of every targeted website.

The findings complicate OpenAI’s account that, to its knowledge, frontier AI discovered the vulnerabilities before attackers did. Earlier discovery by researchers can give defenders preparation time, but Amazon’s observations show that this period did not remain free of suspected hostile activity through public disclosure.

OpenAI’s proposed response framework covers validating flaws, preparing fixes, giving relevant organizations advance notice, coordinating with governments and supporting deployment after publication. Maintainers would retain responsibility for deciding who receives information before disclosure.

The business challenge is broader than finding a bug. The European statement warns that commercial and open-source products can include PHP without making that dependency obvious to their users. Identifying which supplier operates the affected software can therefore be part of the response, alongside installing an update.

Temporary protections also have limits. PHP’s maintainers say measures that reduce exposure to particular attack paths do not make an installation safe; updating provides the complete solution. WordPress similarly advises users to confirm their PHP update rather than relying solely on a host’s firewall preparations.