OpenAI 发现 PHP「世纪漏洞」，全球 70% 网站都危险了
「PHP 是世界上最好的语言」……但 10 月 21 日夜晚，却成了所有 PHP 运维人员的噩梦之夜。
两天前，代号 PHP2Shell 的两个 PHP 高危远程执行漏洞信息意外泄露，表明在最危险的情况下，攻击者可以通过 PHP 的漏洞，在服务器执行任意代码。
漏洞「偷跑」，也倒逼 PHP 基金会在昨晚 10 点提前发布正在筹备的紧急安全补丁。
你可能听说过但不熟悉 PHP，但只需要知道，全世界约有七成网站后台都使用了这项技术，使用比例高达 92% 的日本，更是成了「重灾区」。
于是，从昨晚开始，无数网站运维人员经历了夺命连环 call，高市早苗亲自发文敦促网站管理者打补丁，半个欧洲的政府与学术机构紧急「拔网线」，二次元熟悉的 niconico、pixiv，乃至《赛马娘》《碧蓝幻想》和任天堂相关游戏服务也纷纷下线，开始了归期未定的漫长升级。
向左滑动查看更多内容
甚至连寿司郎，也宣布小程序暂停运行，恐怕要到明天下午，我们才能重新用上小程序预约。
一场安全风暴，正在席卷全球。
PHP 出问题，全球互联网抖一抖
和很多类似的技术一样，PHP 是一种「不出圈」但相当重要的互联网基础设施。
想象你登录一个电商网站，下单买了一本书。这个过程中你看见的只是前端的网页，但服务器其实在后台处理了诸如验证密码、检查库存、查询订单、生成网页等动态功能。
在相当多网站上，实现这些后台工作的编程语言就是 PHP。
根据统计，目前有 69.4% 的网站后端用到了 PHP，WordPress、Drupal、Joomla 等流行的网站管理系统都建立在 PHP 之上，维基百科使用的 MediaWiki，同样主要由 PHP 编写。
但就像曾经种满全球的大麦克香蕉被黄叶病团灭一样，普遍性本身也是一种脆弱性：如果 PHP 在底层上出现安全问题，可想而知，几乎全球的网站都会受到牵连。
这刚好就是我们目前面临的状况。
这次爆出的两个漏洞，对应编号分别为 CVE-2026-109618 和 CVE-2026-109619，PHP 基金会给它们定下的严重程度评分达到 9.2 分（满分 10 分）。
能拿到「高分」，自然是有相应的危险程度。
举个不完全准确的比喻：假设 PHP 是麦当劳中央厨房，后厨工作人员会根据显示屏上的下单情况来制作餐食。而最新曝光的内存管理漏洞，就相当于让黑客有机会修改显示屏内容，让后厨给自己做出源源不断的「自助餐」。
这里的「自助餐」内容，可以是让服务器乖乖交出存储的用户资料，也可以是修改新闻网站的内容，或者将电商网站的收款信息改成黑客的账户；进一步地，往服务器植入勒索软件或以此为跳板入侵更大的系统也不在话下。
只要条件合适，这些 RCE（远程代码执行）漏洞，几乎可以让黑客为所欲为。
这也是为什么，在安装好补丁之前，用到 PHP 的网站与服务宁愿彻底下线，不给攻击者留下可乘之机。
但官方发布补丁，也不代表万事大吉。一方面，不是所有网站管理者都有时间、有意愿去更新系统；另一方面，基建上的惯性导致目前世界上还有近半网站使用的是已经结束了支持周期的旧版 PHP，而这些旧版分支从一开始就得不到补丁。
也就是说，如果不彻底升级 PHP 版本，这些网站从今天开始几乎都是处于「裸奔」状态。
早在补丁推送之前，思科、WatchTowr 和 360 等公司的安全团队已经根据公开信息复现了攻击流程，而 Cloudflare 也开始观察到小规模的攻击行为。
可以预想，在主要公司密锣紧鼓地升级的同时，还有很多潜在的攻击者正在搜寻那些怠于维护的网站，随时准备分食服务器上潜藏的资源。
竞速大逃杀，现在才刚刚开始。
AI 开始发掘「沉睡漏洞」
值得注意的是，根据官方复盘，这起事件中 AI 扮演了相当核心的角色。
引发这次安全风暴的漏洞，其实早在 2024 年就被发现——
当时，PHP 基金会启动了一次外部安全审计，而在检查审计结果时，PHP 维护团队的成员开始注意到，在内存不足的情况下 PHP 会存在安全隐患。但由于没有找到稳定可复现的攻击方式，加上修复核心代码可能会影响兼容性，修复工作一拖再拖，始终没有提上日程。
但 AI 的出现改变了一切。
今年 6 月，OpenAI 发起了一项名为「Patch the Planet」（修复星球）的安全项目，这个颇有「万物皆可 Patch」气质的计划找来了多家安全公司，合作利用 AI 发现并报告开源项目的漏洞，作为开源项目的 PHP 自然也被列入研究目标。
在 AI 的助力下，不到两个月时间，合作安全公司 Calif 研究人员就找到了利用 2024 年那个漏洞执行远程攻击的方式。
几天后，另一支研究团队 Wiz Research 也借助 Wiz Atlas、Gemini 4 Argon 和 GPT-5.6 Sol 几个模型，独立发现了相同的漏洞。
根据 Wiz 的公告，PHP 这两个漏洞堪称「软件历史上最严重的漏洞之一」，但利用这些漏洞的复杂程度「几乎超出了人类研究人员能够实现的范围」。
换句话说，PHP 这辆「老爷车」之所以能拖着严重的底层缺陷如常运转，很大程度上要归功于人类攻击者的能力不足。但前沿 AI 赋予的智能，让原本不可能的攻击成为了现实。
好在，这一次安全团队并非毫无准备——
早在 8 月发现可行的攻击方式之后，PHP 技术团队就将修复漏洞放到了最高优先级，联手 OpenAI 和多家安全公司组成的联盟秘密开发了相关补丁，因此才能在漏洞信息偷跑后第一时间释出补丁。
但很显然，没有理由总是「正义联盟」胜出。OpenAI 在官方博客中也提醒：
尽管这些漏洞影响重大，但它们只是我们已经发现并正在着手修复的多个重要漏洞中的一部分，而我们预计，未来几个月类似发现的数量将大幅增加。
计算机系统比想象中更脆弱
PHP2Shell 并不是第一次系统基建出现问题，进而影响大量互不相关的服务的事件。
从写入计算机教科书的「千年虫」事件，到 2014 年的 Heartbleed 漏洞、2021 年的 Log4Shell 远程代码执行漏洞以及 2024 年的 CrowdStrike 更新导致大规模蓝屏，每隔一段时间，现实都会用一种略带荒诞的崩溃事件提醒我们：
计算机系统的基础，远比我们想象中更脆弱。
而脆弱的根源，很大程度上又可以归结于一点，亦即：
现代计算机系统的复杂程度，远远超过了它在底层技术上的多样性。
我们熟悉的网站和 app，形式上可能相差甚远，但剥离表层，大量服务实际上依赖同一批编程语言、软件库、操作系统和网络协议。
维护这些基建设施的，却往往只是一些摇摇欲坠的开源基金会。
以 PHP 基金会为例，这个为全世界 70% 网站提供技术底座的基金会，2025 年全年捐助收入共 73 万美元，只有 11 名全职及兼职签约开发者。他们贡献了当年 PHP 核心代码库约 42% 的提交。
如果不寻找其他收入来源，仅靠这个项目能得到的收入，其实是低于美国软件开发者收入中位数的。
客观来说，这些事务繁忙却收入不高的资深开发者就算直接开摆，恐怕也没有人能苛责他们，更何况他们实际上甘愿用爱发电、尽职尽责。
但即使维护团队有意愿去维护，他们也未必能够得偿所愿：PHP 已经太受欢迎，以至于围绕它形成了庞大的软件生态，任何核心组件的改动都可能导致大量第三方插件与扩展失效。
于是，整个系统只能继续冒着蒸汽、掉着煤渣，以一种「你不问我不说」的姿态继续前行。
某种意义上，整个互联网世界，就像《Viva la vida》的歌词所说：
And I discovered that my castles stand upon pillars of salt and pillars of sand
在过去的文章里，很多读者朋友会对 AI 厂商大肆渲染的安全叙事不置可否，更倾向于把它们归入到营销炒作的范畴。但只要意识到我们生活的地基有多么脆弱，恐怕就不会那么轻视 AI 带来的安全风险了——
真正的危险不在于 AI，而在于我们年久失修的计算机系统基建。
PHP2Shell 事件最终会造成多大损失还有待时间给出答案，但随着 AI 让越来越多沉睡多年的缺陷重新暴露出来，这样的紧急修复，恐怕也不会是最后一次。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Open AI discovered the PHP's century loophole.
"PHP is the best language in the world." But on the night of October 21, it became night of nightmares for all PHP operators.
Two days ago, two PHPs, code-named PHP2Shell, were accidentally leaking information about high-risk remote execution loopholes, indicating that, in the most dangerous circumstances, the attackers could enforce any code on the server through a PHP loophole.
The leak "stealing" also forced the PHP Foundation to prepare an emergency security patch at publish at 10 pm last night.
You may have heard of PHP but you're not familiar with PHP, but all you need to know is that about 70 percent of the world's web sites use this technology, using up to 92 percent of Japan, and it's becoming a "hard-hit zone."
As a result, since last night, numerous website operators have experienced a life-threatening chain of events, calling, the city of Gao’s own message urging website managers to fill the patch, the emergency “drill line” between half the European governments and academic institutions, and the second-tier niconico, pixiv, as well as the Soma’s Blue Fantasy and Nintendo-related game services, which have begun a long and undetermined escalation.
Slide left to see more content
Even Sushiro has suspended the small program, and I'm afraid we won't be able to use the small program again until tomorrow afternoon.
A security storm is sweeping the globe.
PHP is in trouble. The global Internet is shaking.
Like many similar technologies, PHP is a "no circle" but rather important Internet infrastructure.
Imagine you're getting a book on an electrician website. All you see in this process are front-end web pages, but servers actually handle dynamic functions like authentication of passwords, checking of inventories, searching for orders, creating web pages, etc.
On a considerable number of websites, the programming language for these back-office tasks is PHP.
According to statistics, 69.4 per cent of the website backends are currently used by PHP, popular website management systems such as WordPress, Drupal and Joomla are based on PHP, which is also produced mainly by PHP.
But just as once the big Mac bananas that were grown around the globe were wiped out by yellow leaf disease, universality itself is a vulnerability: if PHP has a security problem at the bottom, it is conceivable that almost all the world's websites will be implicated.
That is exactly what we are facing.
The two loopholes, CVE-2026-109618 and CVE-2026-109619, respectively, were rated by the PHP Foundation at 9.2 points (10 points).
It's a natural hazard to get a high score.
Take an inaccuracy metaphor: Assuming that PHP is McDonald's Central Kitchen, the late cooks make meals based on the list on the screen. The most recently exposed memory management gap is the opportunity for hackers to modify their screen content so that they can cook for themselves a continuous "self-help".
The content of the "basket" here can be either to get the server to hand over stored user information, or to modify the content of the news site, or to change the collection information from the electrician website to a hacker's account; and further, to plant blackmail software on the server or use it as a springboard to hack larger systems.
As long as the conditions are right, these RCE gaps can almost make hackers do whatever they want.
That is why, before the patch is installed, the PHP website and service prefers to be completely offline and leave no room for the attackers.
But the official publish patches do not mean anything. On the one hand, not all website managers have the time and willingness to update the system; on the other hand, the inertia in infrastructure has led to the fact that nearly half of the world ' s web sites now use the old version of the PHP that has ended the support cycle, which has not been patched from the outset.
That is to say, without a complete upgrade of the PHP version, these sites are almost all "naked" from today.
Long before the patch was sent, the security teams of Cisco, WatchTowr and 360 had reproduced the attack on the basis of public information, while Cloudflare had begun to observe small-scale attacks.
It can be envisaged that, while major companies are escalating, many potential assailants are searching for sites that are poorly maintained and ready to share the resources hidden on servers.
The race to flee has just begun.
AI, start digging the sleeping hole.
It is worth noting that according to the official review, AI played a very central role in this incident.
The leak that triggered this security storm was actually discovered in 2024...
At that time, the PHP Foundation initiated an external security audit, and at the time of the audit, members of the PHP maintenance team began to notice that there was a security risk to PHP in the event of inadequate memory. However, the failure to find a stable and re-emergible pattern of attack, coupled with the potential impact of the restoration of the core code on compatibility, has been delayed and has remained on schedule.
But the emergence of AI changed everything.
In June this year, Open AI launched a security project called "Patch the Planet," a project with a "Patch for All" mass that brought together security companies to work with AI to identify and report on the gaps in open-source projects, which are naturally included as an open-source project.
With the help of AI, in less than two months, the cooperative security company Calif researchers found a way to use the 2024 loophole to carry out a remote attack.
A few days later, another research team, Wiz Research, found the same gaps independently, using several models from Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol.
According to Wiz’s bulletin, these two loopholes are described as “one of the most serious gaps in software history”, but the complexity of using them is “almost beyond the reach of human researchers”.
In other words, the PHP, which has been able to drag down a serious bottom flaw, as usual, is largely attributable to the lack of capacity of human attackers. But front-line AI gives intelligence that makes impossible attacks a reality.
Fortunately, the security team is not unprepared this time...
As early as August, when a viable attack was discovered, the PHP technical team placed repairing the loophole at the top of its priorities, and a coalition of Open AI and various security companies secretly developed the patches to release them first after the leaking information.
But clearly, there is no reason why the Justice League should always win. Open AI also reminds the official blog that:
Despite the significant impact of these loopholes, they are only part of a number of important gaps that we have identified and are in the process of repairing, and we expect a significant increase in the number of similar discoveries in the coming months.
Computer systems are more fragile than they thought.
PHP2Shell is not the first time that a system infrastructure problem has arisen, thus affecting a large number of unrelated service events.
From the "Millennium Bug" event, which was written in computer textbooks, to the 2014 Heartbled loophole, the 2021 Log4Shell remote code implementation loophole, and the 2024 CrowdStrike update, which led to large-scale blue screens, reality reminds us every once in a while with a slightly absurd collapse:
Computer systems are much more fragile than we thought.
The root causes of vulnerability can be largely attributed to:
Modern computer systems are far more complex than their technological diversity at the bottom.
Our familiar web sites and apps may differ considerably in form, but they are detached from the surface, and a large number of services actually rely on the same programming language, software repository, operating systems and network protocols.
The maintenance of these infrastructure facilities is often limited to volatile open-source foundations.
In the case of the PHP Foundation, which provides a technical base for 70 per cent of the world ' s web sites, the total annual contribution revenue in 2025 was $730,000, with only 11 full-time and part-time contractors. They contributed about 42% of the year's PHP core code library.
If alternative sources of income are not sought, the revenue from the project alone is actually lower than the median income of software developers in the United States.
To be objective, senior developers who are busy and poorly paid for these services, even if they do so directly, are afraid that no one will be able to punish them, not to mention the fact that they are willing to use love to generate electricity and do their duty.
But even if the maintenance teams are willing to do so, they may not be able to get what they want: PHP has become too popular to have a large software ecology around it, and any change in the core component could lead to a large number of third-party plug-ins and extensions not working.
So the whole system can only continue to steam, drop the slag, and move forward in a "you don't ask me not to say."
In a sense, the entire Internet world, as Viva la vida says:
And I discovered that my castles stand upon pillars of salt and pillars of sand
In the past, many readers and friends have been able to ignore the security narrative that has been widely publicized by AI manufacturers, preferring to place them in the realm of marketing. But as long as we realize how fragile the foundations of our lives are, I'm afraid that the security risks of AI will not be underestimated.
The real danger is not in AI, but in the infrastructure of our old computer systems.
The PHP2Shell event will ultimately have time to answer the question of how much damage it will cause, but this emergency repair will probably not be the last as AI rediscovers more and more sleeping defects.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.