不用登录、不用密码，黑客一个请求，就能远程接管全球70%的网站服务器！这两天，全球7成网页的基座PHP，被曝存在2个非常严重的漏洞。团队随之与Open AI一起，秘密组织了一场修复。不过，让人脊背发凉的是：这场本该保密两个月的抢修，最后48小时却崩盘了。
凤凰网科技综合多方信息发现，网上率先有匿名帖先放出漏洞的 CVE 编号，并得到了PHP 官方的确认；之后，安全公司 Calif 误发了攻击演示视频，视频随之被删但副本在全网扩散开；最后，软件供应链安全公司Chainguard 提前发布了含修复的构建镜像，等于把漏洞位置提前递到了攻击者手里。
补丁还没正式落地，第一波攻击已经上路。而OpenAI甚至预告：这次爆出来的，只是开胃菜。
藏在 PHP 里 14 年的漏洞，被GPT翻了出来
“PHP曝出‘php2shell’远程代码执行漏洞(CVE-2026-109618、CVE-2026-109619)。如您的业务使用PHP，请尽快自查并升级至安全版本，避免服务器遭受攻击。”
今天，不少负责运维、安全、开发的同学，都收到了上面这则安全风险通告，提醒使用PHP的用户尽快升级安全版本。
首先得说，PHP 可不是没人管的冷门工具。它平时藏在服务器后台，负责把数据库里的文章、商品、用户信息拼成网页，再端给浏览器：WordPress、Drupal、Magento 这些撑起大量网站的产品，底层都靠它运行。
二十多年来，无数开发者、安全研究员反复研究过它的代码，一遍遍做审计、打补丁。一个被如此盯防的软件，怎么还能在眼皮底下藏一个 14 年的雷？
并且，这两个漏洞的评分都是 9.2，属于“严重”级别。攻击者在特定条件下无需登录，只需要发送一个请求，就能在业务代码运行前直接拿到服务器执行权限。等于大楼的门禁还没起作用，后门已经开了。
听起来像科幻，但这就是这次事件的真实情况。
而帮忙挖出这个漏洞的是GPT-5.6 Sol等人工智能。OpenAI 在公告中确认，安全公司 Calif 使用 GPT-5.6 Sol 发现了 PHP 的关键漏洞；Wiz Research 也独立发现并报告了相关问题，并在研究中使用了 Wiz Atlas、Gemini 4 Argon 和 GPT-5.6 Sol。
图｜源自Open AI官方发帖
具体来说，在研究员的带领下，GPT-5.6 Sol 等AI审核了PHP 的底层代码，发现PHP存在“数据只写了一半，却被当成完整数据”的异常路径。接着，把这条线索整理成可复用的审计规则，继续往下查，又发现了第二个问题。
这件事的意义不止于“AI 找到了一个 bug”。PHP 已经运行了十几年，很多代码在当年是够用的，只是按今天的安全标准重新看有些过时。AI 把这种沉睡风险加速翻了出来，也让防御方抢在攻击者之前获得了一个时间窗口。
但问题也随之而来：如果 AI 能帮好人更快地发现漏洞，它是不是也能帮坏人更快地利用漏洞？
更棘手的是，这次漏洞的细节在补丁发布之前就已经外泄，攻击者甚至跑到了补丁前面。 而 PHP 官方已发布的几个修复版本，PHP 8.1 及更早版本都不能获得正式更新，只能靠源码补丁或发行版回补。
也就是说，很多网站的困境不是修不修？而是能不能修？修到哪一步？修完会不会把业务搞崩？
漏洞被发现，只是第一关。真正难的是后面：资产清点、版本兼容、测试验证、灰度升级、重启进程、确认旧进程是否还在跑。AI 把发现漏洞的速度拉满了，但修复这件事，仍然卡在组织、流程和旧系统里。
确实，在AI的加持下，挖漏洞的门槛正在降低，但防御的速度却未必跟得上。AI 让“发现”变得廉价，可“修复”仍然昂贵。
AI 能更快找到漏洞，互联网为什么还是修不好？
其实，这两处漏洞并不是今天才被发现。
PHP 基金会表示，CVE-2026-109618 的线索，最早出现在 2024 年 PHP 项目的一次外部安全审计中。当时，一名核心开发者在修复审计发现的问题时，已经注意到了相关隐患……“迟迟没有修复，是一次结构性失败”。
PHP 基金会的说法，揭示了一个困境：对程序员来说，找出漏洞往往比修好它更简单。
就像PHP基金会自己说的，多种修复方案都各有代价：有的可能破坏向后兼容，有的可能引入新的拒绝服务风险，还有的需要协调大量第三方扩展。
修复工作因此停了又启、启了又停，问题被搁置了下来。直到这一次，它被GPT-5.6 Sol发现。
其实，更大的矛盾是，全球不少商业巨头、创业公司，都在免费享用 PHP 这类开源软件打下的地基，可地基的维护，长期压在一小批志愿者和捉襟见肘的资金上。
PHP 基金会曾多次公开呼吁：那些靠开源赚得盆满钵满的公司，有没有拿出与自己依赖程度相匹配的投入？
去年9月，Java、Python、PHP、Rust、Eclipse 等基金会还联合 OpenSSF 发布声明，直接喊话大企业：“开源基础设施并非免费”，长期免费使用不可持续，商业用户应该为基础设施付费。
说白了，问题不是没人看见漏洞，而是看见之后，谁来付钱、谁来承担兼容风险、谁来为那些早已无人维护的系统兜底？
AI 的出现，把这种问题放大得更明显了。它让“发现漏洞”变得前所未有地快，却无法自动解决“修复漏洞”这件最费人力的事。补丁要不要打？打了会不会出事？由谁来审？企业要不要立刻更新？老系统迁不迁移？……每一步都需要组织和人来做决策。
AI 确实提高了找到问题的速度，却把修问题的责任，原样留给了本来就已经人手不足的生态。
所以，当 AI 越来越擅长找漏洞，互联网真正的短板到底在哪？
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。
We can take over 70% of the global web server remotely without login, without password, without hacking! On these two days, the PHP, the cornerstone of 70 pages worldwide, was exposed to two very serious loopholes. And the team, together with Open AI, secretly organized a restoration. What cooled the back, however, was that the last 48 hours of the two-month operation, which was supposed to be confidential, collapsed.
The phoenix technology synthesis found that the Internet first had a CVE number with a leak, which was officially confirmed by the PHP; that, after that, the security company Calif had mistook a video of the attack, the video was deleted and a copy spreads across the Internet; and that, finally, the software supply chain security company Chaingular had pre-empted a restored building mirror with publish, which would have delivered the leak position to the attackers in advance.
The patch hasn't officially landed yet, and the first wave is on its way. And Open AI even predicted that the explosion was just an appetizer.
The 14-year-old loophole hidden in PHP was turned over by GPT.
“PHP exposed a `php2shell' remote code implementation gap (CVE-2026-109618, CVE-2026-109619). If you use PHP for your business, check yourself and upgrade to a secure version as soon as possible to avoid attacks on the server.”
Today, many of the students in charge of transport, security and development have received the above-mentioned security risk circular, reminding users who use PHP to upgrade the security version as soon as possible.
First of all, PHP is not an unmanageable tool. It is normally hidden in the back of the server, and is responsible for collating articles, commodities, user information in the database into a web page and re-routing it to browsers: WordPress, Drupal, Magento, which runs the base of a large number of web-based products.
For more than two decades, its code has been repeatedly studied by numerous developers and security researchers, auditing and patching. How can you hide a 14-year-old thunder under your nose?
Moreover, the ratings for both loopholes are 9.2, which is “serious”. The attackers were not required to log in under certain conditions, but were required to send a request to obtain the server to execute permission directly prior to the operation of the business code. It's the same as the building's doorblock. The back door is open.
That sounds like science fiction, but that's the real story.
And it's GPT-5.6 Sol and so on. In its bulletin, Open AI confirmed that the security company Calif used GPT-5.6 Sol to identify key PHP loopholes; Wiz Research also independently identified and reported relevant issues and used Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol in its research.
It's from Open AI official post.
Specifically, under the leadership of the fellows, GPS-5.6 Sol and others reviewed PHP's bottom code and found that PHP had an abnormal path of “data only half-written but considered complete data”. The lead was then collated into reusable audit rules, which continued to be checked downwards and a second problem was identified.
This is more than "AI found a bug." PHP has been in operation for more than a decade, and many of the codes were useful in the year, but they are somewhat outdated in terms of today ' s safety standards. AI accelerated this sleeping risk and allowed the defensive to gain a window of time before the attackers.
But the question follows: if AI can help good people to find the hole more quickly, can it help bad people to exploit it more quickly?
Even more problematic is the fact that the details of this loophole were leaked before the patch publish, and the attackers even ran to the patch. The PHP official has several restored versions of publish, and neither PHP 8.1 nor earlier versions can be formally updated, but can only be replaced by source patches or distributions.
In other words, isn't there a problem with many websites? But can you fix it? Which one? Will it ruin the business?
The hole was found, just the first level. The real difficulties lie behind: asset counts, version compatibility, test validation, greyscale upgrades, restarting the process and confirming whether the old process is still running. AI filled the speed at which the loopholes were discovered, but repairing the matter was still stuck in the organization, processes and old systems.
Indeed, under AI, the threshold for opening holes is being lowered, but the pace of defence is not necessarily kept pace. AI makes “discovery” cheap and “repairable” still expensive.
AI, you can find a loophole faster. Why can't you fix the Internet?
Indeed, these two loopholes were not discovered today.
According to the PHP Foundation, the CVE-2026-109618 trail was first presented in an external security audit of the PHP project in 2024. At that time, a core developer, in repairing the problems identified by the audit, had noted the associated pitfalls.
According to the PHP Foundation, a dilemma is revealed: for programmers it is often easier to find a loophole than to fix it.
As the PHP Foundation itself has said, there are costs associated with a variety of rehabilitation programmes: some have the potential to undermine backward compatibility, others have the potential to introduce a new risk of denial of services, and others need to coordinate large third-party expansion.
As a result, the work was halted, rekindled and suspended. Until this time, it was discovered by GPT-5.6 Sol.
Indeed, the greater paradox is that many business giants and start-ups around the world are free of charge using the foundation of open-source software such as PHP, which can be maintained over a long period of time on a small group of volunteers and overstretched funds.
The PHP Foundation has made several public appeals: have companies that earn a lot from open sources come up with inputs commensurate with their level of dependence?
In September last year, foundations such as Java, Python, PHP, Rust, Eclipse and others also joined OpenSSF publish in stating that it was not sustainable to speak directly of large enterprises: “open-source infrastructure is not free” and that commercial users should be paid.
To put it simply, the question is not who sees the holes, but who pays, who takes the risks of compatibility and who runs the bottom of systems that are no longer maintained.
The emergence of AI has magnified this problem. It makes the “discovery of the loophole” unprecedented, but it does not automatically address the most labour-intensive task of “repairing the loophole”. You want a patch? Is something wrong? Who's going to review it? Should the company update it? Is the old system moving?... each step requires organizations and people to make decisions.
AI did increase the speed of finding the problem, leaving the responsibility for repairing it to the already understaffed ecology.
So, when AI gets better and better at finding holes, where is the real Internet panel?
Auto-fetched and translated by AI Atlas. All rights belong to the original author.