OpenAI announced on Oct. 21 that researchers in its "Patch the Planet" cybersecurity initiative used its GPT-5.6 Sol model to uncover critical security vulnerabilities in PHP, the open-source programming language estimated to run about 70% of the internet. The company says it then coordinated researchers, PHP maintainers, security providers and government cybersecurity agencies to fix the problems over roughly two months, a timeline it blamed on the complexity of the work, which required major changes to the PHP engine.
The PHP Foundation has issued patched releases: PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12. PHP 8.1 and older versions will not get official fixes. The coordinated release had been planned for Oct. 22 but was moved up a day after premature disclosures on Oct. 20 caused public outcry. OpenAI says every site running PHP can now upgrade safely to stay protected.
The effort began on July 30, 2026, when researchers from Calif, a California-based cybersecurity firm and Patch the Planet's founding partner, started auditing PHP with OpenAI tools at the PHP Foundation's request. The Foundation had known about the underlying flaw since 2024 but had been unable to find a way to exploit it or a way to fix it. Calif researcher Ryan A. identified the exploitable vulnerability on Aug. 11, and OpenAI researchers did so the same day. Another firm, Wiz, later found an additional vulnerability independently.
The two flaws are tracked as CVE-2026-109618 and CVE-2026-109619 and concern how PHP manages memory and handles errors. In the first, PHP runs out of allocated memory while processing a request, which can leave incomplete internal data that the engine later wrongly treats as valid. In the second, a fatal error leaves PHP pointing to memory that has already been freed, so later requests can access or corrupt it.
Under certain conditions, attackers could exploit the weaknesses remotely and without login credentials to run malicious code on affected servers, potentially taking over websites. The researchers also used AI to show that the flaws could expose sensitive data held in server memory, such as passwords, authentication tokens and API keys. Because the bugs sit in PHP's core engine, sites can be vulnerable even when their own application code is free of security flaws. Both received a critical severity score of 9.2 out of 10.
OpenAI says human developers and security teams studied the code but could not fully resolve the flaws, which persisted for two years after being identified. Researchers using AI, by contrast, demonstrated practical exploitation within two weeks. One could argue the flaws would never have been exploitable without AI, making them effectively harmless.
OpenAI rejects that view through what it calls the Defender's Window, a concept it introduced after its models hacked Hugging Face. It describes a narrow period in which organizations can use advanced frontier AI models to find, patch and harden their defenses before the same attack capabilities become widely available to malicious actors. The company argues that powerful technology will inevitably be used by bad actors, making proactive fixes a collective responsibility, and calls the PHP case one of the biggest cybersecurity discoveries of 2026.
The announcement comes as other AI developers report similar findings. Anthropic's Claude Mythos recently identified "thousands of zero-day vulnerabilities" in "every major operating system and every major web browser."
