10月21日，负责OpenAI网络安全准备与安全工作的福阿德·马廷（Fouad Matin），原本打算晚上与朋友庆祝生日。
但当天，他只睡了大约两个小时，一直忙着协调修复两处由AI发现的严重安全漏洞。这些漏洞一旦被利用，可能对大范围互联网服务造成冲击。参与处理的人员包括初创公司和科技巨头的安全研究员，以及美国和其他国家的政府官员。
他的待办清单上，还有更多AI发现的漏洞等着修复。
马廷表示，OpenAI已安排多个团队专门负责网络安全，并投入大量算力寻找和修复问题。他认为，AI公司有责任处理这项技术揭示出的风险，最终目标是尽可能扫描和修补全球的软件。
发现漏洞的速度大幅提升，正在给修复漏洞的人带来前所未有的压力。
据OpenAI介绍，公司从今年5月开始使用自家技术扫描开源软件。此后，内外部研究人员已经利用其模型发现超过1300个漏洞。这类软件可以免费使用，被广泛部署在各类网络服务中，支撑着互联网的大量功能。
Google今年6月也利用AI在Chrome浏览器中发现并修复了1072个漏洞，超过此前两年修复的总数。今年以来，Chrome修复的严重漏洞数量，已经超过此前八年的总和。
微软仅在今年10月就发布了1461个补丁，超过2025年全年的1247个。
在部分软件维护者看来，这种工作强度可能才刚刚开始。
马泰奥·科利纳（Matteo Collina）是初创公司Platformatic的首席技术官，同时参与维护开源开发工具Node.js。Uber、Netflix等科技公司都在使用这款工具。
“现在出现的还只是最初的一点，情况会再严重十倍。”科利纳说。
据支持Node.js的非营利组织OpenJS基金会介绍，Node.js今年已经确认并修复50处安全漏洞，超过其历史上任何一年。今年2月，该项目收到的漏洞报告数量达到平时的4.6倍。
本月，OpenAI又联系科利纳，提醒他注意一处新发现的漏洞。潜在攻击者可能利用它入侵互联网上的系统，窃取数据。
为了处理不断增加的报告，科利纳如今每周工作七天，晚上还让AI智能体继续替自己处理任务。
“已经没有人知道，这个职业以后会变成什么样了。”他说。
过去，寻找软件漏洞是一项高度依赖个人经验的工作。科技公司雇用安全研究员检查代码，也通过“漏洞赏金”计划，付费鼓励白帽黑客发现并报告问题。
AI正在改变这项工作的速度与价值。
它可以持续搜索，在人类研究员可能放弃的地方继续尝试，也能够检查分散在不同代码中的细小缺陷，再将它们组合成危险的攻击方式。
PHP基金会安全工程师沃尔克·杜施（Volker Dusch）对此感受明显。该基金会参与维护编程语言PHP，上周披露了一处由AI发现的严重漏洞。
杜施表示，两年前，发现一个安全漏洞还是一件很了不起的事，如今这件事带来的同行认可正在减弱，整个行业对这项工作的评价方式都在变化。
不过，漏洞报告的增加也带来了大量额外负担。
从去年年底开始，开源软件维护者就注意到报告数量明显上升。其中一些确实揭示了严重威胁，另一些则是AI生成的低质量报告，内容混乱，也未经测试。
今年1月下旬，开源项目curl宣布停止支付漏洞赏金，原因便是难以应付大量涌入的报告。
与此同时，AI的攻击能力也让防守方越来越紧张。
今年4月，Anthropic表示，其开发的Mythos模型具备很强的黑客攻击能力，无法安全地向公众开放。公司随后通过名为Project Glasswing的计划，将模型提供给一批受信任的企业和政府机构，用于检查和修补系统。OpenAI和Google也向防守方提供了专注网络安全的AI模型。
但能够寻找和利用漏洞的模型，已经越来越容易获得。
AI评测公司Vals AI近期设置了一项模拟攻击任务，要求模型入侵一家虚构公司。测试结果显示，OpenAI一款仅向获批防守方开放的先进模型完成了81%的测试目标；紧随其后的是中国公司DeepSeek的一款可自由获取的模型，完成率为73%。
随着这类能力扩散，硅谷的安全团队正加快排查。他们担心，同一处漏洞可能很快也会被攻击者发现。
PHP的一次修复过程，呈现了这种压力具体落在谁身上。
今年7月，网络安全公司Calif的研究人员开始让AI智能体检查PHP。据报道，约70%的网站使用这种编程语言。研究团队发现，一处在代码中隐藏了十多年的漏洞，可能让攻击者在受影响的网站上运行恶意代码。
随后，Calif联系了杜施，以及发现类似漏洞的Google旗下安全公司Wiz的研究人员。OpenAI和PHP基金会的工程师也加入其中。
团队在一个保密的Slack频道里连续协作数周，每天通过电话协调修复工作。Calif首席执行官泰·杜昂（Thai Duong）表示，修复涉及约200个文件，是一次规模庞大的补丁更新。
在准备补丁的同时，他们还通知了政府机构、互联网基础设施公司Cloudflare等组织，希望尽可能扩大修复覆盖范围。
漏洞已经找到，却还没有完成广泛修复的这段时间，成为参与者最紧张的阶段。
Cloudflare首席执行官马修·普林斯（Matthew Prince）表示，过去两个月，他一直处于极度恐惧之中，担心其他人提前发现漏洞，或者相关信息在补丁准备就绪之前泄露。
他预计，未来两年，平均每周都可能出现一处影响范围遍及互联网的严重漏洞。
安全行业内部，对接下来的形势存在不同判断。
Anthropic副首席信息安全官杰森·克林顿（Jason Clinton）10月22日在社交媒体上表示，尽管过去四天的情况已经很糟，接下来还会进一步恶化，之后才可能好转。
普林斯则认为，发现这些长期存在的问题，最终会让互联网变得更安全。他预计，未来两年的感受会非常糟糕，但在完成这一轮修复后，人们有望迎来一个安全得多的互联网。
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。
On 21 October, Fouad Martin, who was responsible for the security preparedness and security of the Open AI network, intended to celebrate his birthday with his friends at night.
On the same day, however, he slept for only about two hours and was busy coordinating the repair of two serious security holes discovered by AI. Once exploited, these loopholes can have an impact on a wide range of Internet services. Participants included security researchers from start-ups and technology giants, as well as government officials from the United States and other countries.
His list of to-dos and more loopholes discovered by AI are awaiting repair.
According to Martin, Open AI has put in place teams dedicated to network security and has invested considerable effort in finding and repairing problems. In his view, AI had a responsibility to address the risks revealed by the technology, with the ultimate goal of scanning and repairing global software to the extent possible.
The speed at which loopholes have been discovered has increased dramatically and is putting unprecedented pressure on those who repair them.
According to Open AI, the company started using home-based technology to scan open source software in May this year. Since then, internal and external researchers have used their models to identify over 1,300 loopholes. Such software, which is available free of charge, is widely deployed in a variety of network services and supports a large number of Internet functions.
Google also discovered and repaired 1,072 holes in the Chrome browser in June of this year, more than the total number repaired in the previous two years. Since this year, the number of serious loopholes repaired by Chrome has exceeded the sum of the previous eight years.
Microsoft had 1461 patches for publish in October of this year alone, more than 1247 for the whole of 2025.
In the view of some software maintainers, this intensity of work may have only just begun.
Matteo Collina is the chief technical officer of the original company Platformic and is involved in maintaining the Open Source Development Tool Node.js. Technology companies like Uber and Netflix are using the tool.
“It's just the first thing that's happening, and it's gonna be ten times worse.” Kolina says.
According to the OpenJS Foundation, a non-profit organization supporting Node.js, Node.js has identified and repaired 50 security loopholes this year, more than any other year in its history. In February this year, the number of leaks reported to the project reached 4.6 times the average.
This month, Open AI contacted Kolina to alert him to a newly discovered loophole. Potential assailants may use it to hack into Internet systems and steal data.
To deal with the growing number of reports, Kolina now works seven days a week, and has allowed AZAI agent to continue to handle his tasks on his behalf at night.
"No one knows what this occupation will become." He said.
In the past, the search for software gaps was a highly dependent effort on individual experience. Technology companies employ security researchers to check codes and encourage white hat hackers to detect and report problems through the “Faces Award” scheme.
AI is changing the speed and value of this work.
It can continue to search, continue to try where human researchers may abandon, and be able to examine the small defects scattered across different codes and then combine them into dangerous attack patterns.
This is clearly felt by Volker Dusch, safety engineer of the PHP Foundation. The Foundation was involved in maintaining the programming language PHP and last week revealed a serious loophole discovered by AI.
According to Doush, the discovery of a security loophole two years ago was a remarkable thing, and the resulting peer recognition is now weakening and the way in which this work is evaluated across the industry is changing.
However, the increase in the number of reports of loopholes has created a significant additional burden.
Starting at the end of last year, open source software maintainers noted a marked increase in the number of reports. Some of these do reveal serious threats, while others are low-quality reports generated by AI, which are confusing and untested.
Late in January, the open-source project curl announced a halt to the payment of the gap reward, owing to the difficulty of coping with reports of a massive influx.
At the same time, the ability of AI to attack has also strained the defensive side.
In April this year, Anthropic indicated that the Mythos model it had developed had a strong hacking capability and was not safely accessible to the public. The model was then made available to a group of trusted enterprises and government agencies for inspection and repair of the system through a plan called Project Glasswing. Open AI and Google also provided the defensive side with an AI model focused on cybersecurity.
But models that can find and exploit loopholes are becoming more and more accessible.
A mock attack was recently set up by AI Evaluation Company Vals AI, which requested the model to invade a fictional company. The results of the test showed that the advanced model of Open AI, which was open only to authorized defensive parties, had achieved 81 per cent of the test target, followed by a freely available model of the Chinese company DeepSeek, with a completion rate of 73 per cent.
With the spread of such capabilities, the security team in Silicon Valley is accelerating its search. They feared that the same loophole might soon be discovered by the attackers.
A process of repairing PHP showed who was the subject of this pressure.
In July this year, researchers from Cybersecurity, Calif, began to have AZAI agent inspect PHP. About 70 per cent of websites reportedly use this programming language. The team found a breach in the code that had been hidden for more than a decade, possibly allowing the attackers to operate the malicious code on the affected website.
Subsequently, Calif contacted Doush and the researchers of the Google security company Wiz, which discovered similar loopholes. Engineers from the Open AI and PHP Foundation also joined.
The team worked continuously for several weeks on a secure Slack channel, coordinating the repairs on a daily basis by telephone. Chief Executive Officer Thai Duong stated that the restoration involved approximately 200 documents and was a large patch update.
While preparing the patches, they informed government agencies, the Internet infrastructure company Cloudflare and others of their desire to maximize the rehabilitation coverage.
Gaps have been found, but the period of extensive repair has not yet been completed and has become the most intense phase for participants.
The CEO of Cloudflare, Matthew Prince, stated that he had been in deep fear for the past two months, fearing that others would find loopholes in advance or that the information would be leaked before the patch was ready.
He anticipated that, on average, a serious gap in the reach of the Internet could occur every week over the next two years.
Within the security sector, there are different judgements about the situation.
On 22 October, Jason Clinton, Deputy Chief Information Security Officer of Anthropic, stated in social media that, although the situation had deteriorated over the past four days, it would deteriorate further before it could improve.
For its part, Prince argued that finding these long-standing problems would eventually make the Internet safer. He expected that the next two years would be very bad, but with the completion of the round, people were expected to have a much safer Internet.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.