Forget the AI Slowdown—the Vulnerability Explosion Is Already Happening
Forget the AI Slowdown—the Vulnerability Explosion Is Already Happening
别管什么 AI 发展放缓了——漏洞大爆发已经来临
Welcome to the inaugural edition of Kernel Panic! A weekly newsletter by Lily Hay Newman and Matt Burgess from inside the new world of privacy and digital security. To receive this newsletter in your inbox each week, sign up here. 欢迎阅读《内核恐慌》（Kernel Panic）创刊号！这是一份由 Lily Hay Newman 和 Matt Burgess 共同撰写的每周通讯，带你深入隐私与数字安全的新世界。若想每周在收件箱中接收此通讯，请点击此处订阅。
AI doomers have recently traded one worst-case scenario for another, putting aside a potential software vulnerability apocalypse to focus on the possibility of rogue AI causing mass human death in the next decade. As AI leaders consider a cooperative slowdown on frontier model development, though, one aspect of the cybersecurity sea change has already arrived thanks to existing, broadly available capabilities in mainstream AI products, including open weight models. AI 的末日论者最近将一种“最坏情况”换成了另一种：他们暂时搁置了对软件漏洞引发灾难的担忧，转而关注未来十年内失控 AI 可能导致人类大规模死亡的可能性。然而，正当 AI 领军人物考虑通过合作来放缓前沿模型开发时，网络安全领域的一场巨变已经因为主流 AI 产品（包括开放权重模型）中现有的、广泛可用的功能而悄然降临。
A tidal wave of vulnerabilities uncovered using AI has only accelerated in recent months—piling more pressure on under-resourced, and very human, IT and security teams and straining volunteers who maintain crucial open source software. Researchers found and disclosed a vast array of vulnerabilities before the rise of AI-enhanced bug hunting as well, but the recent surge is clear. 近几个月来，利用 AI 发现的漏洞如潮水般涌现，且势头有增无减。这给资源匮乏的 IT 和安全团队（由人类组成）带来了更大的压力，也让维护关键开源软件的志愿者们不堪重负。虽然在 AI 辅助漏洞挖掘兴起之前，研究人员也曾发现并披露过大量漏洞，但近期的激增态势显而易见。
Microsoft said last week that it has issued patches for 974 CVEs so far this month, setting a new record. (CVEs, or common vulnerabilities and exposures, is cybersecurity jargon for confirmed software flaws.) In July, Oracle shipped 1,448 patches compared to 309 in July 2025. Google Chrome’s two major version releases in June included 1,072 patches, more than all of the vulnerability fixes shipped in the prior 23 big releases combined. And Mozilla said in April that it found 271 vulnerabilities in Firefox during one bug hunting sprint using Anthropic’s Mythos model. 微软上周表示，本月迄今已发布了 974 个 CVE 的补丁，创下新纪录。（CVE，即“通用漏洞披露”，是网络安全领域对已确认软件缺陷的术语。）今年 7 月，甲骨文（Oracle）发布了 1,448 个补丁，而 2025 年 7 月仅为 309 个。谷歌 Chrome 浏览器 6 月份的两个主要版本更新包含了 1,072 个补丁，超过了此前 23 次重大版本更新中漏洞修复数量的总和。Mozilla 也在 4 月份表示，在使用 Anthropic 的 Mythos 模型进行的一次漏洞挖掘冲刺中，他们发现了 Firefox 中的 271 个漏洞。
Across the board, there have been a stunning 66,401 CVEs recorded as of Wednesday this week, according to Jerry Gamblin, the head of research at Empirical Security and founder of RogoLabs, which runs the CVE analysis project cve.icu. By September 16 last year, cve.icu had logged a total of 33,512 CVEs—almost half the current total. For all of 2022, the year OpenAI launched its first version of ChatGPT, cve.icu recorded 25,000 CVEs. 根据 Empirical Security 研究主管、CVE 分析项目 cve.icu 的创始人 Jerry Gamblin 的数据，截至本周三，全球已记录的 CVE 总数达到了惊人的 66,401 个。去年 9 月 16 日时，cve.icu 记录的总数为 33,512 个，仅为当前总数的一半左右。而在 OpenAI 发布第一版 ChatGPT 的 2022 年全年，cve.icu 记录的 CVE 总数为 25,000 个。
Among both security and AI researchers, experts have been divided about whether this spike and other impacts of AI on cybersecurity will be catastrophic or instead magnify existing dynamics and challenges. Some have pointed out that slow patch adoption and lagging investment in cybersecurity broadly already gave attackers many advantages that led to hacking disasters before the rise of AI. But as vulnerability discovery numbers have continued to rise, and the discussion has become less theoretical, the two sides have seemed to move a bit closer. 在安全和 AI 研究人员中，专家们对于这种激增以及 AI 对网络安全的其他影响究竟是灾难性的，还是仅仅放大了现有的动态和挑战，一直存在分歧。一些人指出，在 AI 兴起之前，补丁更新缓慢和网络安全投资滞后就已经让攻击者占据了许多优势，从而导致了黑客灾难。但随着漏洞发现数量持续上升，讨论也变得不再那么理论化，双方的观点似乎正在趋于一致。
“I don’t think it’s overblown,” Gamblin says of the apparent explosion in vulnerability findings across the industry. “What I would push back on is the idea that a bigger number is itself the harm. More CVEs is not more vulnerability. It’s more known vulnerability, which is mostly the system working.” “我不认为这是夸大其词，”Gamblin 在谈到整个行业漏洞发现的明显激增时说道。“我想要反驳的是那种认为‘数字越大本身就是危害’的观点。更多的 CVE 并不意味着更多的漏洞，而是意味着更多的‘已知漏洞’，这在很大程度上说明系统正在发挥作用。”
The fear, though, is that vast vulnerability discovery will mean developers getting outpaced on patching, software users who can’t patch fast enough, and an array of escalating cyberattacks fueled by more attackers discovering novel vulnerabilities on their own using AI. As Britain’s National Cyber Security Center puts it, “Just finding vulnerabilities does nothing to improve your security.” 然而，令人担忧的是，海量的漏洞发现意味着开发人员在修补速度上将落后于攻击者，软件用户无法及时更新补丁，以及更多攻击者利用 AI 自行发现新型漏洞，从而引发一系列不断升级的网络攻击。正如英国国家网络安全中心所言：“仅仅发现漏洞并不能改善你的安全性。”
For now, many researchers tell us that there is at least a tenuous balance between AI accelerating bug discovery and AI aiding defenders. “Actors, just like industry, are trying to figure out, ‘where do I use AI?’” says Matthew Olney, director of threat intelligence at Cisco Systems. 目前，许多研究人员告诉我们，在 AI 加速漏洞发现与 AI 辅助防御者之间，至少存在一种脆弱的平衡。思科系统威胁情报总监 Matthew Olney 表示：“攻击者和行业一样，都在试图弄清楚：‘我该在哪里使用 AI？’”
As the situation continues to evolve, an AI slowdown of whatever form—be it regulation or an industry accord—could perhaps/hopefully prevent AI from carrying out a mass human extermination event, but it cannot stop the vulnerability tsunami that has already arrived as a result of existing AI tools. 随着局势的不断演变，无论以何种形式（无论是监管还是行业协议）放缓 AI 发展，或许（希望）能防止 AI 实施大规模的人类灭绝事件，但它无法阻止由现有 AI 工具所引发的、已经到来的漏洞海啸。
As RogoLabs Gamblin puts it, “Discovery scales with compute. Remediation scales with people—and people are the part you can’t buy more of in a quarter.” 正如 RogoLabs 的 Gamblin 所言：“漏洞发现随着算力而扩展，而漏洞修复则随着人力而扩展——而人力，正是你无法在一个季度内通过金钱买到更多的那部分资源。”