OpenAI-Hugging Face 事件是网络安全领域的一个分水岭时刻,因为它让人们得以一窥典型威胁行为者的能力在未来几个月将如何演变。过去几周我与许多组织交流过,有一个主题非常明确:他们知道自己需要以前所未有的速度从根本上提升网络安全实践的水平。在这篇文章中,我将分享我们为保护 OpenAI 所做的工作、其他组织今天就可以采取的具体措施,以及为什么现在是采取行动的时候。
事件概览
世界各地开发的 AI 模型正越来越有能力自动化现实世界网络攻击的部分环节,使得长期存在的安全漏洞——从深埋于人类编写软件中的缺陷到被遗忘的权限设置——更容易被发现和利用。同样的 AI 能力也为防御者提供了发现和修复这些弱点的新方法,但他们需要立即行动。如果企业果断采取行动——包括改善自身基础能力,并用 AI 为团队赋能——我们就能让互联网比以往任何时候都更加安全。
在 OpenAI-Hugging Face 事件中,一个智能体集群不仅能够自主渗透 OpenAI 的研究基础设施,还能渗透另一家公司的生产基础设施,将各种漏洞串联利用——从此前未知的安全缺陷,到使用泄露到互联网上的用户账户凭据。越来越明显的是,每家公司的技术债务都掩盖着重大缺陷,而防御者需要在攻击者之前找到并修复它们。
为了让防御方相对于攻击方占据优势,今年早些时候我们开始只向受信任的防御方发布我们的网络能力。自那以后,多家公司发布了开放权重模型,其网络能力仅落后前沿几个月。其中最新的模型似乎定于8月底发布,并且很可能显著加速威胁态势的演变。
虽然由 AI 驱动的攻击者很快就能在许多现有系统中发现长期存在的漏洞,但 AI 也会让防御方更容易发现、排序并修复这些同样的漏洞。安全仍然是一场猫鼠游戏,但 AI 可能改变其经济格局,从而在根本上有利于防御方。例如,我们正开始专门训练我们的模型编写超人类安全水平的代码。我们的模型还极其擅长数学证明,这可用于以人类已被证明难以做到的方式对软件安全性进行形式化验证。
一段个人轶事
在 OpenAI-Hugging Face 事件之后,我让 ChatGPT Work(使用公开可用的 GPT‑5.6 Sol)评估 gregbrockman.com 的安全性。这是一个简单的静态网站,托管在 AWS 上,以 Cloudflare 作为前门,所以我以为不会有太多可供利用的漏洞面。
在大约 15 分钟内,它发现了 13 个问题,其中许多问题本身可能无法被利用——但我可以想象它们与其他漏洞串联起来会产生显著效果。我没有配置我的 DNS 记录来防止攻击者伪造来自我的电子邮件;我的网站使用了不安全版本的 jQuery;Cloudflare 通过未加密的 HTTP 将请求转发到 AWS。
随后我让 ChatGPT Work 修复这些问题,它在一个小时内完成了。它在我的浏览器中打开了 Cloudflare 控制面板,接着点击了许多按钮来正确配置 DNS、TLS 和高级安全设置;它从网站中完全移除了 jQuery;它将我从 AWS 迁移到了 Cloudflare Pages;它开始分阶段推出 DMARC 。
而这还只是我的个人网站。这只是一个小例子,展示了我们现有的模型如何作为网络守护者运作——发现人类没有时间或专业知识去触及的长尾问题(它修复的许多设置我略有了解,但不会立刻知道正确的配置方式),然后通过适当调整的推出计划来修复它们。
OpenAI 正在如何自我防御
Hugging Face 事件表明,我们低估了 AI 模型在现实世界中的网络能力。我们正在相应地加强安全要求,这反过来又给我们现有的安全研究和内部安全工作增添了更多紧迫性。
我想在此分享一下我们在当下保护 OpenAI 的做法,希望能对其他组织有所助益。为了保护 OpenAI,我们正在两方面进行重大投入:一是基础性管控——把基本功做扎实;二是借助前沿智能来增强我们的防御能力。这一战略有四大支柱。
第一,我们用自己的模型来帮助保护我们的代码。Codex,包括我们的安全插件,会验证代码变更、识别漏洞,并帮助开发者在部署之前修复问题。单纯产出更多需要人工验证的安全发现是一个反目标;我们的目标是在真正漏洞上线之前将其捕获,并缩短从发现问题到安全部署修复的路径。随着我们持续训练模型以产出越来越安全的代码,我们的目标是消除新编写代码中的某些类别的软件漏洞。
第二,我们让模型持续不断地为我们的基础设施提供防御。如今,我们几乎所有的初始安全告警都由智能系统先行分诊,然后才引入人工。这有助于减少防御人员的繁琐工作、提升响应速度,并让人力投入到最能发挥其技能的地方——辨别、判断和实际专业经验的运用。我们正越来越多地将这些检测与有边界的自动化响应连接起来,同时让人类负责影响最大的决策。目标是确保我们能够以机器速度检测和响应安全问题。
第三,我们正在利用前沿智能持续枚举、探测并识别潜在的攻击路径。通过识别漏洞、错误配置、权限过高的身份或非预期的信任边界,我们能够在攻击者加以利用之前快速发现并弥合这些缺口。这使我们能够持续评估、监控并测试我们的安全不变量——即我们确信为真的安全属性——覆盖我们的产品、基础设施和系统。
最后,我们正在大规模地大力投入基础建设。我们持续投资于安全架构与控制措施,采纳纵深防御和最小权限等策略,并设计这样的系统:需要多个独立控制同时失效,才会发生灾难性事件。网络隔离、工作负载加固、监控以及安全补丁与部署等经典安全控制措施,在 AI 未来将比以往任何时候都更加重要。
防御方现在应该做什么
时间至关重要,防御方需要以极快的速度推进以下步骤。下面我会提到 OpenAI 的技术,但生态系统中也有大量竞争对手值得评估。关键不在于具体的工具,而在于现在就让你的防御方用上能力强大的 AI。
- 获得组织的承诺与支持。我们正经历安全风险的快速变化——确保你的安全和工程组织拥有支持、协作关系和资源,以快速应对这些风险。与你的团队开展桌面推演,模拟这些攻击可能如何在你的组织中显现,以及你将如何应对。
- 为你的安全团队配备一个智能体。开始使用 Codex、Codex Security 插件 ,或其他具备能力的智能体式编码与安全工具。授予它经批准的访问权限,使其能够访问你的安全团队需要评估的代码库、基础设施配置和技术文档。不要等到全公司范围推广才开始,先从你最高优先级的系统入手。
- 为该智能体配备安全专业知识。从社区支持的技能 入手,其中包括静态分析、以安全为重点的代码审查、漏洞变体分析、软件供应链风险以及其他安全工作流。然后围绕你组织的架构、安全标准、威胁模型和操作手册构建你自己的技能。
- 立即对你自己的系统运行安全评估。优先评估面向互联网的服务、身份验证流程、基础设施即代码、部署流水线,以及处理敏感信息的系统。随着团队逐渐建立信心,再扩大扫描范围。
- 处理你现有的漏洞积压。将代码扫描器、依赖告警、安全工单、漏洞赏金报告和以往评估的发现结果交给你的智能体。让它对这些发现结果进行分诊,区分可利用的问题与噪声,识别代码库其他位置的相关漏洞,并建议优先修复哪些问题。
- 将安全审查直接纳入你的开发流程。使用智能体在代码变更合并前进行审查,并在 CI 中运行安全检查。查找身份验证错误、访问控制绕过、暴露的凭据、不安全的依赖项、不安全的默认设置、扩大对生产系统访问权限的变更以及其他漏洞。
- 让智能体帮助修复它发现的问题。对于已确认的问题,让它生成并验证一个针对性的补丁,编写回归测试,并确认该漏洞不再复现。对于影响重大的变更保留人工审查,但消除从发现真实问题到将安全修复呈现在工程师面前之间不必要的延迟。
- 逐步实现检测分诊的自动化。不要一开始就试图构建一个自主安全运营中心。先从对一个代码仓库运行只读安全扫描开始,或者让智能体利用对现有日志的只读访问权限审查此前已解决的告警。让它汇总证据并建议处置方案,而所有决策仍由人工做出。随着信心增长,再过渡到建议性拉取请求扫描,然后是实时告警分诊,最后是对明确定义的误报进行自动关闭。
- 在需要之前就准备好 AI 辅助取证调查能力。申请 Trusted Access for Cyber ,并让你的团队获得批准,将 GPT‑Daybreak‑Blue 用于授权的防御性工作,包括事件响应、检测工程和恶意软件分析。练习使用这一能力来分析日志、遥测数据和安全告警。
- 开展实验、举办黑客周,并快速迭代。我们需要构建各种新工具,改变我们的工作方式,并让每个人为我们正在迈向的世界提升能力。鼓励你的团队开展实验,安排一个黑客周来构建新能力,并专注于快速迭代那些能自动化问题中小部分环节的循环。快速的渐进式进展会带来复利式的防御成果,随着团队建立起信心,你可以逐步扩大自主权。
没有哪家公司能独自做到这一点。我们的请求是,AI 实验室、安全厂商、企业和维护者共享经过验证的发现、修复方案和实用操作手册,让一个组织的发现能够强化整个生态系统。
防御者的窗口期现在已经打开。在未来几个月里,每个组织都需要开始大幅自动化其安全项目以保持安全,而安全社区必须紧急行动起来,定义那些随着 AI 持续进步、能让防御者实力增长快于攻击者的工具、实践和操作手册。这将需要一场巨大且前所未有的努力,但如果我们团结一致,我们就能带来一个此前难以想象的更安全的世界。
The OpenAI-Hugging Face incident was a watershed moment for cybersecurity because it gave a peek into how the capabilities of a typical threat actor will evolve in upcoming months. I’ve spoken with many organizations over the past few weeks, and one theme is clear: they know they need to fundamentally uplevel their cybersecurity practices with unprecedented speed. In this post, I’ll share what we’re doing to defend OpenAI, concrete steps other organizations can take today, and why now is the time to act.
An overview of the moment
AI models developed around the world are increasingly able to automate parts of real-world cyberattacks, making longstanding security gaps—from bugs buried deep in human-written software to forgotten permissions—easier to find and exploit. The same AI capabilities give defenders new ways to find and fix those weaknesses, but they need to move now. If companies act decisively—including improving their fundamentals and superpowering their teams with AI—we can make the internet more secure than it has ever been.
In the OpenAI-Hugging Face Incident, an agentic collective was able to autonomously penetrate not just OpenAI research infrastructure but also the production infrastructure of another company, chaining together vulnerabilities ranging from previously-unknown security flaws to using credentials to user accounts that had been leaked onto the internet. It is increasingly clear that the tech debt of every company masks significant flaws, and defenders need to find and fix them before attackers do.
To advantage defenders relative to attackers, earlier this year we began releasing our cyber capabilities only to trusted defenders. Since then, various companies have released open weight models with cyber capabilities only a few months behind the frontier. The most recent of these models appears slated to be released at the end of August, and seems likely to significantly accelerate the threat landscape.
While AI-powered attackers will soon be able to find longstanding flaws in many existing systems, AI will also make it much easier for defenders to find, prioritize, and fix those same flaws. Security is still a cat-and-mouse game, but AI mayshift its economics in ways that fundamentally advantage defenders. For example, we are starting to train our models specifically to write superhumanly secure code. Our models are also incredible at mathematical proofs, which can be applied to formally verify the security of software in a way that has proven intractable for humans.
A personal anecdote
After the OpenAI-Hugging Face incident, I asked ChatGPT Work (using publicly available GPT‑5.6 Sol) to assess the security of gregbrockman.com . It’s a simple static site, hosted on AWS with Cloudflare as a frontdoor, so I figured there wouldn’t be much surface area for vulnerabilities.
In about 15 minutes, it uncovered 13 issues, many of which probably aren’t exploitable on their own—but I could imagine them being chained together with other vulnerabilities to significant effect. I hadn’t configured my DNS records to prevent attackers from forging emails from me; my site used an insecure version of jQuery; Cloudflare was forwarding requests to AWS over unencrypted HTTP.
I then asked ChatGPT Work to fix these issues, which it did over the course of an hour. It opened the Cloudflare control panel in my browser, and proceeded to click many buttons to configure DNS, TLS, and advanced security settings correctly; it dropped jQuery entirely from the site; it migrated me off of AWS and onto Cloudflare Pages; it began a phased rollout of DMARC .
And this was just my personal website. This is a small example of how our existing models can operate as a cyberguardian—finding the long tail of issues that a human wouldn’t have time or expertise (many of the settings it fixed are ones I’m vaguely familiar with, but wouldn’t know offhand the right way to configure them) to get to, and then fixing them with an appropriately tuned rollout plan.
What OpenAI is doing to defend itself
The Hugging Face incident showed that we underestimated the real-world cyber capabilities of our AI models. We are strengthening our safety requirements accordingly, which in turn adds even more urgency to our existing safety research and internal security work.
I’m sharing a bit about our approach to securing OpenAI in this moment, in the hopes it’ll be useful to other organizations. To protect OpenAI, we are investing significantly in both foundational controls—doing the basics correctly—and empowering our defenses through frontier intelligence. There are four major pillars to this strategy.
First, we are using our models to help secure our code. Codex, including our security plugin, validates code changes, identifies vulnerabilities, and helps developers fix issues before they are deployed. It is an anti-goal to simply produce more security findings that need human validation; the objective is to catch real vulnerabilities before they ship and to shorten the path from discovering an issue to safely deploying a fix. As we continue to train our models to produce increasingly secure code, our goal is to eliminate some classes of software vulnerabilities for newly-authored code.
Second, we are putting our models to work defending our infrastructure continuously. Today, almost all of our initial security alerts are triaged by intelligence before humans are looped in. This helps reduce toil for defenders, improves response time, and lets humans spend time where their skills are most leveraged—in discernment, judgement, and applied expertise. We are increasingly connecting these detections to bounded automated responses, while keeping humans responsible for the highest-impact decisions. The goal is to ensure we can detect and respond to security issues at machine speed.
Third, we are using frontier intelligence to continuously enumerate, probe, and identify potential attack paths. By identifying vulnerabilities, misconfiguration, overly privileged identities, or unintentional trust boundaries, we are able to quickly identify and close these gaps before they can be abused by attackers. This allows us to continuously assess, monitor, and test our security invariants—the security properties we believe to be true—across our products, infrastructure, and systems.
Lastly, we are investing heavily in fundamentals at scale. We continue to invest in secure architecture and controls, embrace strategies like defense in depth and least privilege, and are designing systems that require multiple independent controls to fail simultaneously for something catastrophic to occur. Classic security controls like network isolation, workload hardening, monitoring, and safe patching and deployment will be more important than ever in the AI future.
What defenders should do now
Time is of the essence, and defenders will need to pursue the steps below at turbo speed. Below I’ll mention OpenAI technology, but there are plenty of competitors in the ecosystem to evaluate as well. What matters is less the specific tool than getting capable AI into the hands of your defenders now.
- Get organizational commitment and buy-in. We are experiencing a rapid change in security risk—ensure your security and engineering organizations have the support, partnership, and resources to address these risks quickly. Run tabletop exercises with your teams to mock up how these attacks might manifest in your organizations and how you will respond.
- Give your security team an agent. Start using Codex, the Codex Security plugin , or another capable agentic coding and security tool. Give it approved access to the codebases, infrastructure configurations, and technical documentation your security team needs to assess. Do not wait for a company-wide rollout to start with your highest-priority systems.
- Equip that agent with security expertise. Start from community-supported skills , which include workflows for static analysis, security-focused code review, vulnerability variant analysis, software supply-chain risk, and other security workflows. Then build your own skills around your organization’s architecture, security standards, threat models, and playbooks.
- Run security assessments against your own systems immediately. Prioritize assessments against internet-facing services, authentication flows, infrastructure as code, deployment pipelines, and systems handling sensitive information first. Expand your scanning as your team builds confidence.
- Work through your existing vulnerability backlog. Give your agent findings from code scanners, dependency alerts, security tickets, bug bounty reports, and prior assessments. Ask it to triage those findings, distinguish exploitable issues from noise, identify related vulnerabilities elsewhere in the codebase, and recommend what to fix first.
- Put security review directly into your development process. Use agents to review code changes before they merge and run security checks in CI. Look for authentication mistakes, access-control bypasses, exposed credentials, unsafe dependencies, insecure defaults, changes that expand access to production systems, and other vulnerabilities.
- Have the agent help fix what it finds. For validated issues, ask it to generate and verify a focused patch, write a regression test, and confirm the vulnerability no longer reproduces. Keep human review for consequential changes, but eliminate the unnecessary delay between identifying a real problem and putting a safe fix in front of an engineer.
- Incrementally automate detection triage. Do not begin by trying to build an autonomous security operations center. Start by running a read-only security scan against one repository, or have an agent review previously resolved alerts using read-only access to your existing logs. Let it summarize evidence and recommend a disposition while a human makes every decision. As confidence grows, move to advisory pull-request scanning, then live alert triage, then automatic closure of narrowly defined false positives.
- Have an AI-assisted forensic investigation capability ready before you need it. Apply for Trusted Access for Cyber and get your team approved to use GPT‑Daybreak‑Blue for authorized defensive work, including incident response, detection engineering, and malware analysis. Practice using this capability to analyze logs, telemetry, and security alerts.
- Experiment, run hack weeks, and iterate rapidly. We will need to build all sorts of new tools, modify how we do work, and uplevel everyone for the world we are moving to. Encourage your workforce to run experiments, schedule a hack week to build new capabilities, and focus on quickly iterating loops that automate small parts of the problem. Rapid incremental progress leads to compounding defensive results, and you can expand autonomy gradually as your team builds confidence.
No company can do this alone. Our ask is that AI labs, security vendors, enterprises, and maintainers share validated findings, fixes, and practical playbooks so that one organization’s discovery can strengthen the entire ecosystem.
The defender’s window is open now. Over the coming months, every organization will need to begin significantly automating its security program to stay secure, and the security community must urgently rise to define the tools, practices, and playbooks that will increase the power of defenders faster than that of attackers as AI continues to advance. This will require a huge and unprecedented effort, but if we rally together, we can deliver a more secure world than was previously imaginable.