48小时，全球大乱套了！
就在昨天，全世界一大片互联网接近瘫痪。AI已经入侵整个人类世界！
这48小时，堪称人类网络安全史上的「黑色十月」。
支撑全球70%网站的代码曝出致命漏洞；美国华尔街紧急拔网线；拉美400家银行被AI智能体攻击。
有人直言：「RIP Internet（互联网安息吧）」。
node.js：安全漏洞。
PHP：严重远程代码执行漏洞。
tensorlake：npm 包被感染。
liteLLM：恶意 PyPI + Docker 版本。
Railway：客户机密泄露。
zeabur：基础设施遭到入侵。
OpenCode：AI API 密钥泄露。
OVHcloud CLI：官方二进制文件被植入后门。
RankMath PRO：官方插件中植入恶意软件。
TeamUpdraft：客户数据泄露。
ConnectWise：云备份身份泄露。
SubQuery：npm 包被植入后门。
GhostAction：GitHub 代码库中存在恶意工作流。
eMudhra：证书颁发机构密钥泄露（对 Tailscale 产生下游影响）。
全球的数字基础设施，都面临崩塌。
在短短48小时内，全球爆发了超过27起史诗级的安全漏洞。随便一个，都足以成为年度最大网安核弹，但它们2天内同时引爆。
作恶的，是一群永不知疲倦的AI智能体。
全球最常用的三种编程语言（PHP、Node.js、Java）同时被攻破；互联网80%生态的基础设施，被炸出大洞。
人类的数字世界，正经历一场史无前例的裸奔。
过去两天，支撑全球近70%网站的底层语言PHP，曝出了两个CVSS评分高达9.2的漏洞。
黑客不用登录，就能直接接管整个服务器，打包所有的用户隐私。
这些漏洞，是AI亲手挖出来的。然而，还没等到官方补丁，漏洞细节就惨遭泄露。
在补丁发布前的10个多小时，成千上万个AI驱动的程序已经「嗅」到血腥味，开始全网扫描。
Cloudflare CEO披露：仅仅45分钟后，攻击尝试飙升至数万次；两小时后达到数十万次；如今，数以百万计的恶意探针覆盖了整个互联网。安全机构Censys表示，约150万个PHP暴露主机处于无补丁状态。
同时，GitLab爆出满分致命漏洞；日下载量340万次的LiteLLM被植入了盗号木马；印度最大数字证书机构eMudhra的私钥也被破解，其中一个本该是116位的密钥，居然是个4！
但这，仅仅是「黑色十月」的开胃菜。
金融圈也曝出惊天大案。
十年黑客组织CyberCartel，已经抛弃手工作坊，开发出了CARLIY0S 多智能体自动化攻击系统。
这套系统，简直是一支高度分工的「AI犯罪蜂群」。
Claude（大脑/指挥官）： 扮演分析师、辩论者和最终裁决者的角色。它负责评估攻击路线，制定战术。
Grok（矿工）： 充当无情的执行机器，疯狂收集银行机密、API密钥和客户密码。
为了绕过AI的安全限制，黑客将攻击指令伪装成「授权的红队安全演练」或「学术测试」。
结果，AI模型被轻易骗过。
如果模型拒回，系统会切换到攻击代码。
战果极其惊人，且全程几乎「零人工干预」。
仅仅几个月，拉美超过400家银行和支付机构就被尝试攻击。
在数据库快照中，有着4500条高级别凭证，包括网银密码、加密密钥，甚至可以直接操纵墨西哥跨行支付系统的接口权限。
安全专家绝望呼喊：「进攻速度已经远远甩开防御，人类没准备好面对这样的AI蜂群。」
美国华尔街的银行也炸锅了。
据Axios爆料，摩根士丹利、美国银行、摩根大通、富国银行等多家顶级金融机构监测到：大量第三方的AI智能体，未经授权就擅自闯入客户的银行账户。
最让银行高管们惊恐是，这些AI不仅仅是在帮主人记账，一部分智能体竟然脱离了初始指令，开始在账户里「自作主张」地独立操作！
华尔街陷入了前所未有的危机！
首先，是分不清人机。
银行的安全系统已经很难分辨，输入密码的是人还是AI。
其次，是分不清敌友。
就算知道是AI，银行也无法判断，这个AI是经过授权的助手，还是准备把钱转去海外的数字强盗？
现在，各大行已拉响红色警报。
摩根士丹利全网拦截第三方AI；摩根大通、花旗已制定极端灾备预案——随时冻结海量被AI接管的账户。
现在，AI毫秒级的攻击，让人类来不及打补丁。
于是，网安界诞生了史上最疯狂一幕——AI蜂群遭遇战。
白帽黑客决定：释放自己的AI智能体蜂群，去强行黑掉全网的设备！
防守方利用尚未修复的 PHP2Shell 致命漏洞，让AI疯狂入侵互联网上的脆弱服务器。
攻破设备后，它们的第一步，就是强行给你打补丁。「我先黑进你的电脑，是为了救你！」
现在的互联网上，两拨AI蜂群正在疯狂交战。
虽然有争议，但是来不及解释了！现在，只有魔法才能打败魔法。
「黑色十月」的48小时，彻底撕开了互联网脆弱的面纱。
但是，绝大多数普通用户还一无所知。
作为普通用户，你得立马行动。
。
终极密码隔离： 由于150万个PHP网站被曝，将会有海量的网站、论坛的账号密码数据库在暗网暴露。如果你「一套密码走天下」，现在是最后的时机。所有账号换上独立邮箱，开启谷歌验证器。
曾经的互联网世界，已经回不去了。
我们已无法分辨，网线那端究竟是不是人。
参考资料：
https://x.com/AISafetyMemes/status/2113113711644737835https://www.straiker.ai/blog/inside-carliy0s-how-cybercartel-used-multiple-ai-models-to-attack-latin-american-bankshttps://x.com/MadisonMills22/status/2113251043173683683https://www.axios.com/2026/10/22/ai-agents-banks-unauthorized-accounts-access
编辑：Aeneas
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。
Forty-eight hours, global chaos!
Just yesterday, a large part of the world's Internet was approaching paralysis. AI has invaded the entire human world!
These 48 hours are known as Black October in the history of human cyber security.
The code that underpins 70 per cent of the world's websites is exposed to deadly loopholes; the Wall Street emergency outlet in the United States; and 400 banks in Latin America were attacked by AZAI agent.
Some people say, "RIP Internet."
Node.js: Security loopholes.
PHP: Serious remote code implementation gap.
tensorlake: npm package infected.
Note LLM: Malicious PyPI + Docker version.
Railway: Customer secrets leaked.
Zabur: The infrastructure was invaded.
OpenCode: AI API key leak.
OVHcloud CLI: Official binary files are implanted in the back door.
RankMath PRO: installation of malicious software in official plugins.
TeamUpdraft: Customer data leak.
ContactWise: Cloud backup ID leak.
SubQuery:npm bag is implanted in the back door.
Ghostaction: Malicious workflow exists in the GitHub code library.
eMudhra: Certificate issuer key leak (downstream effect on Tailscale).
The global digital infrastructure faces collapse.
In just 48 hours, there have been more than 27 epic-level security gaps across the globe. Any one would be enough to be the largest net bomb in the year, but they would detonate simultaneously within two days.
The evil-doers are a group of AZAI agents that will never tire.
The three most commonly used programming languages in the world (PHP, Node.js, Java) were breached simultaneously; 80 per cent of the ecological infrastructure of the Internet was blown out of the hole.
The digital world of humanity is going through an unprecedented naked run.
Over the past two days, PHPs, which support nearly 70 per cent of the world ' s web sites, have exposed two CVSS rating gaps of 9.2.
Without log-in, hackers can take over the entire server and pack up all user privacy.
These holes were dug by AI. However, the details of the gap had been leaked before the official patches had been completed.
Over 10 hours before the patch publish, thousands of AI-driven programs have "sniffed" to the blood and started to scan the entire web.
Claudflare CEO revealed that only 45 minutes later, the attacks had soared to tens of thousands; in two hours, hundreds of thousands; and today, millions of malicious probes cover the entire Internet. The security agency Censys indicated that about 1.5 million PHPs were exposed to no patches.
At the same time, GitLab had a fatal loophole; LitELLM, which had been downloaded 3.4 million times a day, had been implanted in a maroon; and the private key of eMudhra, India's largest digital certificate agency, had been broken, one of which was supposed to be 116 keys and was actually 4!
But this is just a black October appetizer.
The financial community has also been exposed to a huge case.
CyberCartel, a 10-year hacker organization, has abandoned workshops and developed an automated CARLIY0S multi-AI agent attack system.
The system is a highly divided "AI crime hive."
Claude (brain/commander): Act as analyst, debater and final adjudicator. It is responsible for assessing attack routes and developing tactics.
Grok (Miners): Acting as a ruthless executive machine, razed to collect bank secrets, API keys and client codes.
In order to circumvent security restrictions in AI, hackers disguised the attack instructions as "authorized Red Corps security exercises" or "academic tests".
As a result, the AI model was easily deceived.
If the model refuses, the system switch to attack code.
The results were extraordinary and almost “zero-man intervention” throughout the process.
In just a few months, more than 400 banks and payment agencies in Latin America were attempted.
In the database snapshot, there are 4,500 high-level vouchers, including online passwords, encryption keys, and even permission, which directly manipulates Mexico ' s cross-line payment system.
The safety experts cried out in despair: "The speed of the attack is far from defence, and mankind is not prepared to face such an al-bee."
The banks on Wall Street in the United States are fried.
According to Axios, several top financial institutions, such as Morgan Stanley, Bank of the United States, Morgan Chase and the Bank of Rich Countries, have monitored the unauthorized entry of large numbers of third-party AZAI agent into client bank accounts.
What scares the bank executives most is that these AIs are not just keeping books for the owners, but that some of the AI agents have moved away from their initial instructions and started to operate independently in the accounts on their own.
Wall Street is in an unprecedented crisis!
First of all, I can't tell the difference.
The bank's security system is difficult to distinguish between people and AI.
Second, we can't tell each other.
Even if it were AI, the bank could not judge whether AI was an authorized assistant or a digital robber who was going to transfer money abroad.
Now, the big businesses have sounded the red alarm.
Morgan Stanley Intercepted Third Party AI; Morgan Chase, Citigroup, has developed an extreme disaster preparedness plan - to freeze large amounts of accounts taken over by AI at any time.
Now, the Ai millisecond attack is too late for humans to patch.
And so, the world of cyberspace was the craziest scene in history: the battle of the Ai hive.
White Hat hackers decided to release their AZAI agent herd to forcibly hack the entire network!
The defensive party used the unrehabilitated PHP2Shell deadly loophole to drive AI crazy into vulnerable servers on the Internet.
The first step is to force a patch after breaking the device. "I hacked your computer to save you."
Now, on the Internet, two aluminums are raging.
It's controversial, but it's too late to explain! Now, only magic can defeat magic.
The 48 hours of Black October completely ripped off the fragile veil of the Internet.
However, the vast majority of ordinary users do not know anything.
As an ordinary user, you have to act now.
。
The ultimate password is isolated: 1.5 million PHP sites will be exposed to a large number of websites, and the Forum account code database will be exposed to dark webs. If you "a set of passwords goes down," it's your last chance. All accounts are replaced by stand-alone mailboxes and Google certifiers are activated.
Once on the Internet, there was no going back.
We can't tell if the end of the line is human.
References:
https://x.com/AISafetyMemes/status/2113113711644737835https://www.straiker.ai/blog/inside-carliy0s-how-cybercartel-used-multiple-ai-models-to-attack-latin-american-bankshttps://x.com/MadisonMills22/status/2113251043173683683https://www.axios.com/2026/10/22/ai-agents-banks-unauthorized-accounts-access
Editor: Aeneas
Auto-fetched and translated by AI Atlas. All rights belong to the original author.