OpenAI Says GPT-5.6 Sol Helped Find PHP Flaw, Warns More Are Coming

OpenAI says its GPT-5.6 Sol model helped a security researcher identify a critical flaw in PHP, the programming language behind a large share of the web. The discovery is part of a wider set of serious vulnerabilities the company says it has found and is working to disclose. Its warning is not simply that AI can find bugs. It is that the volume of credible findings could soon rise faster than maintainers and organisations can prepare fixes.

In a coordinated disclosure post published on 21 October, OpenAI said researcher Calif found one of the PHP vulnerabilities using GPT-5.6 Sol through its Patch the Planet initiative. Security company Wiz independently found and reported both PHP flaws. The distinction matters. This was a coordinated research and repair effort, not a case of one AI system single-handedly discovering and resolving two bugs.

The PHP maintainers released emergency updates on Wednesday for two critical flaws that could allow unauthenticated remote code execution in affected circumstances. The fixes are in PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. The vulnerabilities are tracked as CVE-2026-109618 and CVE-2026-109619, according to the official PHP release notice.

OpenAI credited the PHP team with leading the fixes and release, while Wiz, Cloudflare and OpenAI contributed research and coordination. Administrators should treat the emergency PHP patch guidance we reported earlier as the immediate operational story. Updating supported installations and checking which applications rely on affected PHP versions are more useful than waiting for further AI-security debate.

The broader development is the pace of discovery. OpenAI says these PHP bugs are among several significant vulnerabilities it has identified, with other issues still being fixed. It expects similar findings to increase sharply as capable models become more widely used in security research. That is the company’s forecast, not proof that every software project is suddenly exposed to a known new attack.

AI tools can help researchers inspect large codebases, form hypotheses and test suspicious behaviour. Yet a plausible report is only the beginning. Maintainers still have to reproduce the issue, assess its impact, write and review a fix, coordinate with downstream distributors and allow users enough time to deploy it. A flood of findings could make that human and organisational work the limiting factor.

This is especially important for widely used open-source software. A flaw in a foundational language or library can travel through hosting providers, applications and supply chains. Publishing technical details too early may give attackers a useful map before defenders can move. Delaying too long can also leave users unaware of the risk. The judgement is difficult even without AI accelerating the number of submissions.

OpenAI proposed a five-stage process that begins with validation and repair, then gives affected organisations, government partners, distributors and vendors advance opportunities to prepare before public disclosure. It also pointed to subsidised access to its Daybreak system for selected defenders. Those steps may help, but they do not replace independent verification, maintainer capacity or clear public guidance when a patch is ready.

The PHP case offers a practical test of that approach. One AI-assisted finding and Wiz’s independent discoveries reached maintainers, several parties coordinated, and patched versions were released. It does not establish that attackers were absent or that AI will always favour defenders. OpenAI says the vulnerabilities were found before attackers to its knowledge, a carefully limited claim.

For software teams, the near-term lesson is straightforward. Keep track of the versions you run, apply the PHP updates where relevant and follow primary advisories. The longer-term question is whether the industry’s disclosure machinery can keep pace when both researchers and attackers have more powerful tools for finding the next flaw.