AI finds critical flaws in major web code

Website owners relying on the PHP programming language now have a new reason to upgrade their software immediately. The language powers an estimated 70% of the internet, and two critical security flaws have been discovered that allow attackers to take control of servers without needing login credentials. These vulnerabilities affect the core engine of the software, meaning even sites with perfectly written application code remain at risk if they run unpatched versions.

The discovery highlights a shifting dynamic in cybersecurity where artificial intelligence is being used to find weaknesses that stumped human developers for years. OpenAI announced on October 21 that its GPT-5.6 Sol model helped researchers uncover these issues under its Patch the Planet initiative. The company described the event as one of the most significant cybersecurity finds of 2026, given how ubiquitous PHP is in the digital landscape.

The PHP Foundation released patched versions to address the problem, including releases 8.2.35, 8.3.36, 8.4.27, and 8.5.12. Older versions, specifically 8.1 and earlier, will not receive official fixes. The coordinated release was originally planned for October 22, but the date was moved forward to October 21 after premature disclosures on October 20 caused public concern.

The timeline of the discovery reveals the gap between human and AI capabilities in this specific case. The PHP Foundation had been aware of the underlying flaw since 2024, but human teams could not find a way to exploit it or a fix. Researchers from Calif, a cybersecurity firm and founding partner of the Patch the Planet program, began auditing the code with OpenAI tools on July 30, 2026. On August 11, 2026, Calif researcher Ryan A. identified an exploitable vulnerability, and OpenAI researchers did the same that same day. Another firm, Wiz, independently found a second related flaw.

The two vulnerabilities, tracked as CVE-2026-109618 and CVE-2026-109619, both received a critical severity score of 9.2 out of 10. The first flaw occurs when PHP runs out of allocated memory during a request, potentially leaving behind incomplete data that the engine later treats as valid. The second involves a fatal error that causes PHP to point to memory that has already been freed. This allows subsequent requests to access or corrupt that memory, enabling remote code execution under certain conditions.

Beyond taking control of servers, the AI-assisted researchers demonstrated that these flaws could expose sensitive data stored in server memory. This includes passwords, authentication tokens, and API keys. The ability to extract this data remotely without authentication makes the threat particularly severe for organizations that store sensitive user information.

OpenAI frames this incident through its concept of the Defender's Window. Introduced after its models hacked Hugging Face, this idea suggests there is a narrow timeframe where organizations can use advanced AI to patch defenses before malicious actors gain similar capabilities. The company argues that since powerful technology will inevitably be used by bad actors, it is a collective responsibility to act proactively. Anthropic recently applied this logic when its Claude Mythos model identified thousands of zero-day vulnerabilities in major operating systems and web browsers.

The length of the two-month coordination period to fix the issues was attributed to the complexity of the task, which required major changes to the PHP engine. While some might argue that the flaws were harmless if humans could not exploit them for two years, OpenAI challenges that view. It maintains that the rapid development of working exploits by AI-wielding researchers within two weeks proves the need for proactive defense in an era of advancing machine intelligence.