PHP2Shell Vulnerability: A High-Risk Bug Identified by AI, Why It Causes Concern Across Global Websites

Introduction

On October 21, 2026, PHP official released a patch to fix two high-risk vulnerabilities numbered CVE-2026-109618 and CVE-2026-109619. The security community collectively referred to these vulnerabilities as “PHP2Shell,” as they allowed unauthorized execution of arbitrary code on PHP-running web servers. The news spread to headlines in major media such as BBC, CNN, and CNBC within 24 hours from security blogs, and government websites in multiple countries temporarily suspended self-checks. Unlike previous vulnerability disclosures, OpenAI’s AI security research project “Patch the Planet” and its partners were listed among the core discoverers this time. For the first time, AI played a key role in disclosing vulnerabilities in the foundational software that affects 70% of global websites, an event worth examining more than just the two CVE numbers.

This combination of vulnerabilities drew global attention not only due to their technical severity but also because they occurred in PHP, an internet infrastructure language. PHP is one of the most widely used server-side scripting languages in history, used behind everything from early blogs and forums to today’s large e-commerce platforms and content management systems. When two high-risk vulnerabilities affect such fundamental software, anyone using it cannot remain indifferent. For security researchers, this marks a milestone in AI-assisted vulnerability discovery; for ordinary internet users, it serves as a reminder: the underlying code of the websites we visit every day may be re-examined through new perspectives.

What Happened

The timeline is not complicated. On August 1, 2026, Calif, using OpenAI GPT-5.6 Sol, first detected an anomaly under the “Patch the Planet” project. On August 20, Google’s security company Wiz confirmed that the vulnerability could enable unauthorized remote code execution (RCE) on Drupal. A day later, Wiz also confirmed that WordPress was affected as well. After two months of coordinated disclosure, PHP official released a patch on October 21, with Wiz, OpenAI, and Calif sharing details simultaneously. The next day, major global media reported extensively, and government websites in some countries chose to suspend operations for investigation.

This two-month coordination period was not a mere waiting period but a typical responsible disclosure process. Disclosing parties, affected vendors, and security teams needed to confirm vulnerability details, assess the impact, prepare patches, and try to prevent malicious exploitation before public release. This coordination required extensive human communication and judgment. AI can speed up discovery but cannot replace these organizational efforts.

PHP2Shell consists of two CVEs, with a CVSS score of 9.2/10, combining unauthorized remote code execution with memory security issues. The impact covers all PHP versions that provide external web requests, including common running environments such as PHP-FPM, Apache mod_php, php-cgi, and FrankenPHP. More problematic is that its historical scope dates back at least to PHP 5 in 2004, meaning that over the past two decades, many websites built with PHP may have related code paths.

The danger of RCE vulnerabilities lies in the fact that attackers can execute any command on servers without needing account passwords. This means a well-crafted request could directly become a backdoor on the server. Memory security issues further conceal the exploitation of these vulnerabilities, as conventional input filtering and permission control may not be sufficient to prevent them. The combination of both places PHP2Shell at a nearly maximum threat level.

Approximately 70% of websites worldwide use PHP, which is a significant reason for the widespread attention given to these vulnerabilities. However, “70% of websites affected” and “the vulnerability affects websites using PHP, accounting for about 70% of global websites” are two completely different statements. The former implies that the internet has been widely compromised, while the latter simply indicates the scale of the attack surface. As of the time of disclosure, security agencies have not reported any actual cases of successful exploitation of this vulnerability, but the potential risk is enough to prompt governments in many countries to take proactive measures to check their websites.

This incident also exposes an old problem: vulnerabilities in basic software are often hidden deeply. PHP 5 to PHP 8 have gone through over two decades of version evolution, with a large amount of historical code and compatibility logic accumulated together. Even the most experienced security researchers find it difficult to manually audit all paths in a short time. The introduction of AI has somewhat turned this “searching for a needle in a haystack” task into a scalable scanning process. This does not mean that AI replaces humans, but it certainly changes the efficiency and cost structure of vulnerability research.

Official Narrative Does Not Match

There are several noteworthy details in the official narrative of this incident.

First, the descriptions of the discovery entities are inconsistent. The PHP official patch announcement, Wiz’s technical report, and OpenAI’s blog all participated in the disclosure, but their focuses differ. OpenAI emphasizes that the Patch the Planet project used GPT-5.6 Sol to complete the discovery; Wiz mentions using Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol. It is not entirely clear whether the actual discovery of the vulnerability was completed independently by AI or through human-machine collaboration. These differences in description directly affect the public’s judgment regarding “AI discovering vulnerabilities.”

Second, there is overlap and competition in the narratives of Google and OpenAI. Wiz has been acquired by Google, and its report includes both Gemini and GPT models, while OpenAI highlights the core role of GPT-5.6 Sol. When AI security research capabilities become news highlights, both companies have an incentive to attribute the discovery to their own technologies. Ordinary readers, only looking at the headlines, may simplify a multi-party collaborative vulnerability discovery into a technical victory of one company.

Third, there is tension between “no successful exploitation reported” and “global alert.” The official narrative generally emphasizes the severity of the vulnerabilities: CVSS 9.2, 70% of websites, government websites down, these numbers combined can easily create an atmosphere that a disaster has occurred. But in reality, this is a potential threat that has not yet been confirmed to be widely exploited. Describing potential risks as real crises is a common exaggeration technique in security incident reporting.

Fourth, there is still a gap between the release of patches and their completion. The official release of patches by PHP is just the first step; whether millions of website and server administrators update them in time is what truly determines the scope of impact. Many times in history, the destructiveness of “high-risk vulnerabilities” did not come from the vulnerabilities themselves, but from a large number of unpatched systems. This time is no exception.

These differences in wording may seem like issues with public relations terminology, but they actually relate to how the public understands the real role of AI in network security. If we exaggerate AI-assisted discovery as independent AI discovery, it may lead decision-makers to overestimate technical capabilities and underestimate the core role of human teams in coordination, judgment, and repair.

What’s being said in the comments section

Discussions on Hacker News focus on two directions. Some believe that AI’s participation in vulnerability discovery is a major turning point, meaning that the way basic software is audited for security will be rewritten in the future. Others question whether AI tools merely accelerate work originally done by human researchers, with the real key tasks of vulnerability identification, impact assessment, and patch coordination still relying on human teams. The core of this debate is not whether AI is useful, but to what extent “AI discovery” has changed the distribution of responsibilities.

On social media, ordinary citizens are more concerned about “whether my website will be attacked” and “whether AI can now independently create or repair vulnerabilities.” Some of this concern is exaggerated by media headlines. For example, statements like “AI discovered 70% of the internet affected” confuse the discoverers, the nature of the vulnerabilities, and the scope of impact, making readers easily assume that AI created a attack that covered seventy percent of the internet.

It’s important to distinguish that PHP2Shell is a set of security flaws that need to be fixed, not an exploited attack event. Reporting vulnerabilities as attack events can mislead readers about the real risks.

Some comments also point out that this incident may accelerate governments’ efforts to review the security of critical infrastructure. As one of the infrastructure for global websites, PHP’s security status has long gone beyond technical boundaries and entered the realm of national network security. If government websites undergo audits due to the same vulnerability, both public opinion and politics will demand stricter security standards and accountability mechanisms. The politicization of security issues is sometimes just as worrisome as the technology itself.

Three things for ordinary users

For ordinary users who do not directly maintain servers, this has three practical implications.

First, confirm whether the websites you manage use PHP and follow up on the patch status. Ordinary users cannot directly read server configurations, but they can observe public statements about their websites. If you manage a corporate website, blog, or e-commerce site with PHP at the backend, you should check whether the service provider or technical team has installed patches and pay attention to any actual attack reports in the next few days. If unsure, you can directly ask the hosting service provider or technical负责人.

Second, be wary of online scams that take advantage of this situation. After major security incidents are disclosed, phishing links disguised as “security updates” or “official patches” often appear. PHP official patches should be obtained through php.net or trusted Linux distribution repositories. Do not click on download links from unknown sources in emails or social media. Any “patch” that requires you to enter server passwords or download suspicious attachments should be suspected first.

Third, re-understand the true meaning of “AI discovering vulnerabilities.” The role of AI in this incident is to assist in discovery and analysis, not to carry out autonomous attacks or repairs. Neither glorifying nor demonizing AI helps in understanding the risks. What truly matters remains the coordination of security teams, the timeliness of patching, and users’ updating actions.

Conclusion

What is most noteworthy about the PHP2Shell incident is the series of issues behind the label “AI discovering vulnerabilities”: how ownership of technical discoveries is defined in multi-party collaboration, how potential risks are amplified during dissemination, and how ordinary users can make accurate judgments amidst information noise.

As of October 24, 2026, the incident is still within a 48-hour hot topic window, with patches already released, but the actual repair progress and impact area still require continuous observation. For website operators, timely updates are the only safe response; for bystanders, distinguishing between “vulnerabilities existing” and “vulnerabilities being exploited” is the premise for understanding such news.

This incident will eventually subside, but the questions it raises will not disappear. As AI becomes more deeply involved in basic software security, what we need are not just smarter models, but also clearer disclosure standards, more responsible media reporting, and more discerning readers. When similar vulnerabilities appear next time, I hope we can see the facts immediately, rather than being led by headlines.

Reference Sources

• PHP official patch announcement (2026-10-21)
• Wiz technical report: PHP2Shell vulnerability analysis
• OpenAI blog: Patch the Planet project disclosure instructions
• Google Cloud Security blog: Response instructions
• BBC, CNN, CNBC reports (2026-10-22)
• Hacker News discussions