This is the “most serious” security flaw discovered by AI to date.
On the evening of October 21, Beijing time, some Australian government websites suddenly became inaccessible. In Japan, several national-level games and a large number of e-commerce websites entered emergency maintenance at the same time. Many security authorities around the world, including the U.S. National Security Agency and the European Union Cybersecurity Agency, have issued emergency warnings one after another, and a wave of preventive website shutdowns has begun around the world.
These websites do not share servers. The only thing they have in common is the backend language, which is Hypertext Preprocessor (PHP, full name: PHP: Hypertext Preprocessor). PHP is an open source programming language. About 70% of the world’s websites rely on it to run. It has a history of 20 to 30 years. Soon, OpenAI issued a document stating that in the past two months, its large model GPT-5.6 Sol assisted in the discovery of two critical PHP security vulnerabilities. These two vulnerabilities are collectively known as PHP2Shell by the cybersecurity community.
On October 20, information related to PHP2Shell was leaked. Before the official update of PHP was released, global cyber attacks had begun. Currently, there are no cases of remote code execution and successful intrusion through PHP2Shell. At 10 pm on the 21st, PHP officially released four new versions. For websites, there is no completely effective defense against the above vulnerabilities, and the only safe option is to update. The vulnerability has been bypassed, but the crisis has raised more questions about global cybersecurity.
Picture/Visual China
未戰先亂
PHP has been nicknamed “the best programming language in the world” by the industry for many years. It is often ridiculed. It has long been criticized for its shortcomings in design, performance, security and other aspects. Today, the “gold content” of this sentence is still rising.
This “PHP crisis” is a typical example of chaos before war. On the morning of October 20, a post in the technology community Hacker News claimed that the PHP security team was about to disclose two serious vulnerabilities. The post went viral before being deleted. American cybersecurity company watchTowr later confirmed within hours that the vulnerabilities existed and reproduced one of them. This forced the PHP security team to advance the update release schedule.
Data released by Amazon shows that the earliest attempt to exploit the vulnerability occurred on October 17. The attack came from an anonymous network, suggesting that the attacker may have obtained the details of the vulnerability during the confidentiality period. Cloudflare, a global cloud platform services company, also reported the emergence of early attacks on the day the PHP update was released. After the update was released, attack attempts increased exponentially. Cloudflare CEO Matthew Prince recently told the media that there were thousands of attack attempts in the 10 hours before the vulnerability was made public. The number jumped to tens of thousands within 45 minutes of being published, and by October 22, it had reached millions.
In some modes, the attack can even be triggered before the website’s own business code is run. In reality, an attacker would first make a request, cause a serious error in PHP, and then allow the worker process to proceed the way he wants. In this way, the attacker has the opportunity to control the server or read other data remaining in the memory. In other words, the vulnerability lies in PHP itself, and no matter how standardized the website code is, it cannot stop it.
The discovery of PHP2Shell points directly to long-standing problems with PHP’s memory management and error handling methods. The PHP Foundation published an article on October 21, saying that the reason why these two vulnerabilities are troublesome is that they are at the core of “almost every request”. Any fix will come at a cost, either breaking backward compatibility, introducing new downtime risks, or requiring a large number of third-party plug-ins to make changes.
Although several back-end engineers said that the vulnerability was more complicated to exploit than expected, and there were no confirmed reports of successful intrusions, panic still enveloped the world in a short period of time. Mizuho Bank, Shizuoka Bank, some services of Mitsubishi Group, Japan Airlines and East Japan Railway have suspended some online services. The British government has also offlined 32 online services.
To this end, the PHP Foundation recently directly made accusations of secondary leaks, accusing the software supply chain company Chainguard of releasing images of certain content more than 6 hours before the update was released, which is equivalent to providing the attackers with the details of the repair work and lowering the threshold for reverse analysis of vulnerabilities. Chainguard argued that the move was a helpless move because there were active attacks at the time.
two days and two months
The cost of PHP repair has actually been seen long ago.
In 2024, PHP underwent its first external security audit in years. While fixing one of the issues, a core developer at the PHP Foundation discovered the first signs of the vulnerability mentioned above. At that time, the team had deduced a possible attack method, but could not find a specific path. The repair plan saw back and forth between several ideas. The PHP Foundation recently admitted that this was a “structural failure” caused by inadequate decision-making processes and understaffing of the security team.
The turning point came this year. On July 30, American cybersecurity company Calif began auditing PHP under OpenAI’s “Mending the Planet” program. The plan was released in June this year to deal with the impact of AI vulnerability mining and maintain network security. Just two days later, they found the first entry point that triggered one of the aforementioned vulnerabilities. On August 11, Calif researcher Ryan Austin reported the complete attack chain to the PHP project team. On the same day, OpenAI independently reported the same concerns. Since then, OpenAI has successively reported more attack paths.
PHP core developers searched for an attack path for a long time but could not find it. Another group of people only spent two days with the help of AI, and it took two months to fix the vulnerability.
“This is not just a vulnerability in a certain website building system.” On October 21, Loris DeGiovanni, founder of cloud security company Sysdig, told the media that every PHP application is exposed to threats, including supplier portals, management backends and forgotten marketing websites that the company itself has no record of in the asset list. This points to a more general problem. In addition to websites, PHP is also widely embedded in industrial gateways, building management systems, home routers, and network storage devices. The update cycles of these places are often measured in years. On October 21, the security media Help Net Security cited the judgment of network security researchers in a report that there are too many forgotten and unmaintained applications that will never be patched, and this crisis will “drag out a long tail.”
This sentence raises a deeper issue: When AI makes it cheap to find vulnerabilities, who will bear the cost of fixing them? The PHP Foundation is almost asking for help at the end of the aforementioned article, hoping that companies, host providers, cloud platforms and governments that use PHP will provide sponsorship. The consensus in the industry is that there are too many people looking for loopholes and too few people fixing them.
OpenAI proposed the term “defender window” in its announcement on October 21. AI finds vulnerabilities before attackers do, buying maintainers and security vendors time to prepare. But OpenAI also admitted that this window is conditional. The window only exists when the capabilities of the model in the hands of the defender are significantly ahead of the capabilities generally available to the attacker. Therefore, the number of cases of major vulnerabilities discovered by AI may increase dramatically in the coming months.
How does the PHP crisis affect ordinary people? On October 22, Austin, one of the first people to discover the vulnerability, told the media that exploiting this vulnerability on someone else’s website was essentially like entering the front door. “Now they can enter the rest of the house, roam freely within the network, and carry out ransomware attacks.” Although PHP runs on the server and is not on the personal device, many experts cautioned that whether personal information is affected depends on the long-term impact. If you receive a data leak text message, etc., you should change your personal password in time.
As of press time, access to some websites that were taken offline for precautionary measures has been restored, but there are still a large number of website services around the world that are interrupted or need to be repaired. In addition, there are tens of millions of PHP systems running on small websites and old equipment that have not been touched in ten years. At the end of this year, the early version of PHP 8.2 will stop security maintenance, and there will be a batch of “difficult to repair” machines.
Reporter: Zhou You
(nolan.y.zhou@gmail.com)
Related Reading
- The Japanese government sets standards, and Osaka and other four places compete to become “deputy capitals”2026-10-28
- It is 4,000 years old and has met Lu Yingong, Liu Xie, Guo Moruo… and also me.2026-10-28
- National standards for accompanying patients will be implemented and service providers will be limited to institutions2026-10-28
- Russia and Ukraine have a regional ceasefire in the Kherson region, but fighting has not stopped in many other places2026-10-28
- A wild winter is brewing. Here’s what meteorologists want you to know.2026-10-28
