The PHP vulnerabilities uncovered with help from OpenAI’s GPT-5.6 Sol show how quickly AI can turn a long-known defect into a practical attack path. Two flaws in PHP’s core engine could let unauthenticated attackers run code on affected servers or expose secrets held in memory. The PHP Foundation has issued fixes across four supported release branches.
OpenAI disclosed the research on Oct. 21 through a press release. The work was part of “Patch the Planet,” a cybersecurity initiative that brought together researchers, PHP maintainers, security companies and government cybersecurity agencies. OpenAI said coordination and major changes to the PHP engine made the repair process complex, taking about two months.
Two memory flaws, rated critical
The issues, tracked as CVE-2026-109618 and CVE-2026-109619, concern PHP’s memory management and error handling. In one case, a request that exhausts allocated memory can leave behind incomplete internal data that the engine may later accept as valid. In the other, a fatal error can leave PHP with a pointer to memory that has already been freed, creating an opportunity for later requests to access or corrupt it.
Under certain conditions, either vulnerability can be exploited remotely without login credentials to execute code on a server and potentially take control of a website. The researchers also demonstrated that attackers could use the flaws to retrieve sensitive data from server memory, including passwords, authentication tokens and API keys. Each vulnerability received a severity score of 9.2 out of 10.
Because the weaknesses sit in PHP’s engine, a site may be exposed even when its own application code contains no security flaw. PHP runs an estimated 70% of websites, making the engine level of the bugs significant well beyond any single application or hosting provider.
From known defect to working exploit
The PHP Foundation had been aware of the underlying issue since 2024, but had not found a way to exploit it or determine a fix. At the Foundation’s request, researchers from Calif, a California-based cybersecurity firm and Patch the Planet’s founding partner, began auditing PHP with OpenAI tools on July 30, 2026.
Calif researcher Ryan A. identified an exploitable vulnerability on Aug. 11, and OpenAI researchers found one that day as well. Security firm Wiz later discovered another vulnerability independently. OpenAI says human developers and security teams had reviewed the code without resolving the defects, while AI-assisted researchers demonstrated practical exploitation methods within two weeks.
The Foundation released PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12 with fixes. PHP 8.1 and earlier versions won’t receive official patches. The release was initially planned for Oct. 22, but moved to Oct. 21 after details surfaced prematurely on Oct. 20 and prompted public criticism.
The narrow window for defenders
OpenAI frames the PHP work as an example of its “Defender’s Window”: a limited period when security teams can use advanced AI models to find and fix weaknesses before similar capabilities become broadly available to attackers. The company says the latest PHP releases let sites update to address these flaws.
There’s a harder question behind that framing. If researchers needed AI to produce a working exploit, one could argue the vulnerabilities posed little practical danger before those tools arrived. OpenAI’s response is that newly accessible attack capabilities make earlier weaknesses newly relevant, so defenders should act before malicious users can apply them.
That concern extends beyond PHP. Anthropic has said its Claude Mythos model identified “thousands of zero-day vulnerabilities” across major operating systems and web browsers, as described in its recent findings. For PHP operators, the immediate takeaway is concrete: check the version in use and move to a patched release where one is available.
