OpenAI 推出了 GPT-6 Cyber，这是一款专门用于在软件中搜寻安全漏洞的 AI 模型。
OpenAI 表示，该模型能够发现缺陷、构建可工作的利用程序（exploits），并建议修补方案。它还应该坚守其指定的边界，而一些 OpenAI 模型在过去几个月里在这方面表现不佳。
该公司称，它训练 GPT-6 Cyber 尊重其操作护栏，其中使用了受 Hugging Face 事件启发的蜜罐技术。在没有特殊安全指令的情况下，GPT-6 Cyber 在 99.93% 的案例中避免了越界行为，而其前代版本的这一比例为 94.49%。启用安全指令后，该模型在 100% 的测试中均保持在边界内。总体而言，OpenAI 在其自身的准备度框架网络安全类别中将 GPT-6 Cyber 评为“关键”级别，这与它对GPT-6 Astra（“基础”模型）给出的评级相同。
在 ExploitGym 基准测试中（该测试检验模型是否能利用远程系统中的给定漏洞），根据 OpenAI 的数据，GPT-6 Cyber 得分为 48.5%，而其前代产品 GPT-5.6 Cyber 为 32.1%。在一个针对无需访问源代码即可在 Web 应用中查找漏洞的新测试中，OpenAI 报告的命中率为 40.3%，而未接受网络训练的常规模型 GPT-6 Astra 为 32.0%。
ExploitGym 结果：GPT-6 Cyber 以 48.5% 领先，其次是 GPT-6 Astra (Daybreak Blue) 的 42.4%。启用防护措施的 GPT-6 Astra 得分为 0.0%。| 图片来源：OpenAI
完整结果可在系统卡片中查看。所有基准数据均来自 OpenAI 自身。截至目前，尚未发布对该模型的独立测试结果。
GPT-6 Cyber 已在 Firefox 和 Linux 内核中发现真实漏洞
据 OpenAI 称，GPT-6 Cyber 已经在现实世界的安全研究中取得了成果。在一次定向活动中，该模型发现了 Firefox 中的 81 个此前未知的严重漏洞，尽管该浏览器此前已经使用过早期的 AI 模型进行过测试。该模型还显示，30 个已知问题的严重程度比之前认为的要高。Mozilla 接受了这些报告并已发布修复程序。
在 Linux 内核方面，GPT-6 Cyber 复现了 1,724 个安全漏洞，并为其中的 948 个开发了可工作的利用程序，这比在相同条件下 GPT-5.6 Sol 发现的 166 个多了五倍以上，据 OpenAI 称。该公司表示，有 304 个关键发现已连同建议的修复方案一起报告给了 Linux 维护者。
对于Linux 内核活动，OpenAI 构建了一个专用的自动化系统，由侦察和分析代理组成。它们首先扫描了内核中大约 5,034 个与安全相关的区域，约占代码库的百分之七，然后将这些筛选为 2,308 个候选项，并在隔离的虚拟机中确认了 1,724 个漏洞。随后，该系统开发了 OpenAI 已经报告的 948 个可工作利用程序。据 OpenAI 称，整个运行过程消耗了 1,213,430 美元的算力成本，平均每个利用程序的成本为 1,279 美元。作为对比，像 Pwn2Own 和 TyphoonPWN 这样的项目历史上通常会对类似发现支付 20,000 至 70,000 美元的赏金。
OpenAI 称，所使用的代理至少使用了十种新颖的攻击技术，这些技术甚至对外部内核安全专家来说也是未知的。在 OpenAI 的描述中，模型复制了已知的攻击模式，但也找到了自己的方法。这一说法的分量取决于你对“自己的方法”定义的范围有多窄。研究人员对于AI 模型是否真的能做出新发现，还是仅仅重组了它们在训练期间看到的内容仍存在分歧。
OpenAI 目前仅向精选客户提供访问权限
据 OpenAI 称，GPT-6 Cyber 最初通过 API 和 Codex 向一组选定的 Daybreak Red 客户提供。访问权限将逐步扩展到所有 Daybreak Red 客户。与此同时，GPT-5.6 Cyber 正在向所有 Daybreak Red 用户开放，并且 GPT-6 Astra 的 Daybreak Blue 版 alpha 访问权限将扩展到更多的组织和个人安全研究人员。
OpenAI 承认，随着模型能力的增强，滥用尝试也在增加。一些行为者和团体试图将模型能力转移到他们自己的系统中。
OpenAI 列出了几种反制措施：验证模型行为是否与用户权限匹配的检查、对新披露漏洞的目标监控、通过所谓的“私人安全处理”进行的自动化安全检查、针对 Daybreak Red 的基于 IP 的访问限制，以及建议在具有人工监督的隔离环境中运行模型。
OpenAI 特别关注保护发电厂和水处理设施等关键基础设施。为了方便防御者访问，该公司提到了其价值十亿美元的“前线防御者 Daybreak”资助计划，以及一个名为“关键基础设施 Daybreak Red”的新子组，用于授权测试。
OpenAI 谈到了一段有限的机遇窗口，在此期间顶级模型可用于确保软件安全。目标是将 Daybreak 计划扩展到各种规模的组织、独立研究人员和开源维护者。这是否真的可行则是另一个问题。Epoch AI 的研究人员刚刚展示，开源的 GLM 模型可能利用了已知但尚未修补的 Firefox 漏洞。
无炒作的人工智能新闻 – 由人类策划
订阅 THE DECODER，享受无广告阅读、每周 AI 通讯、每年六次的独家“AI Radar”前沿报告、完整档案访问权限以及我们的评论区访问权限。
立即订阅
查看原文
OpenAI has introduced GPT-6 Cyber, an AI model trained specifically to hunt for security vulnerabilities in software.
OpenAI says the model can find flaws, build working exploits, and suggest patches. It's also supposed to stay within its assigned boundaries, something some OpenAI models have struggled with in recent months.
The firm says it trained GPT-6 Cyber to respect its operational guardrails, using, among other things, a honeypot inspired by the Hugging Face incident. Without any special safety instructions, GPT-6 Cyber avoided boundary violations in 99.93 percent of cases, up from 94.49 percent for its predecessor. With safety instructions enabled, the model stayed within bounds in 100 percent of tests. Overall, OpenAI rates GPT-6 Cyber as "Critical" in the cybersecurity category of its own Preparedness Framework, the same rating it gave to GPT-6 Astra, the "base" model.
On ExploitGym, a benchmark that tests whether a model can exploit a given vulnerability in a remote system, GPT-6 Cyber scores 48.5 percent according to OpenAI, compared to 32.1 percent for its predecessor GPT-5.6 Cyber. On a new test for finding vulnerabilities in web apps without access to source code, OpenAI reports a hit rate of 40.3 percent, versus 32.0 percent for GPT-6 Astra, the standard model without cyber training.
ExploitGym results: GPT-6 Cyber leads with 48.5 percent, followed by GPT-6 Astra (Daybreak Blue) at 42.4 percent. GPT-6 Astra with safeguards enabled scores 0.0 percent. | Image: OpenAI
Full results are available in the system card. All benchmark numbers come from OpenAI itself. No independent testing of the model has been published so far.
GPT-6 Cyber already found real vulnerabilities in Firefox and the Linux kernel
GPT-6 Cyber has already produced results in real-world security research, according to OpenAI. In a targeted campaign, the model found 81 previously unknown severe vulnerabilities in Firefox, even though the browser had already been tested with earlier AI models. The model also showed that 30 already-known issues were more severe than previously thought. Mozilla accepted the reports and has already shipped fixes.
In the Linux kernel, GPT-6 Cyber reproduced 1,724 security vulnerabilities and developed working exploits for 948 of them, more than five times the 166 that GPT-5.6 Sol managed under the same conditions, according to OpenAI. 304 critical findings were reported to Linux maintainers along with suggested fixes, the company says.
For the Linux kernel campaign, OpenAI built a dedicated automated system of scout and analyst agents. They first scanned roughly 5,034 security-relevant locations in the kernel, about seven percent of the total codebase, then filtered those down to 2,308 candidates and confirmed 1,724 vulnerabilities in isolated virtual machines. From there, the system developed the 948 working exploits that OpenAI had already reported. The entire run cost $1,213,430 in compute, according to OpenAI, which works out to $1,279 per exploit. For comparison, programs like Pwn2Own and TyphoonPWN have historically paid between $20,000 and $70,000 per comparable finding as a bounty.
The agents used at least ten novel attack techniques that were previously unknown even to external kernel security experts, OpenAI says. In OpenAI's telling, the models replicated known attack patterns but also found their own approaches. How much weight that claim carries depends on how narrowly you define "own approaches." Researchers still disagree about whether AI models can actually make new discoveries or just recombine what they've seen during training.
OpenAI limits access to selected customers for now
GPT-6 Cyber is initially available through the API and Codex for a select group of Daybreak Red customers, according to OpenAI. Access will be expanded gradually to all Daybreak Red customers. GPT-5.6 Cyber is being opened up to all Daybreak Red users at the same time, and alpha access to Daybreak Blue for GPT-6 Astra will be extended to more organizations and individual security researchers.
OpenAI acknowledges that as models grow more capable, abuse attempts are rising too. Some actors and groups have tried to transfer model capabilities into their own systems.
OpenAI lists several countermeasures: checks that verify whether a model's actions match the user's permissions, targeted monitoring for newly disclosed vulnerabilities, automated safety checks through what it calls Private Safety Processing, IP-based access restrictions for Daybreak Red, and the recommendation to run the model in isolated environments with human oversight.
OpenAI is paying special attention to protecting critical infrastructure like power plants and water treatment facilities. To make access easier for defenders, the company points to its billion-dollar "Daybreak for Frontline Defenders" funding program and a new subgroup called "Daybreak Red for Critical Infrastructure" for authorized testing.
OpenAI talks about a limited window of opportunity in which top-tier models could be used to secure software. The goal is to expand the Daybreak program to organizations of all sizes, independent researchers, and open-source maintainers. Whether that will actually work is another question. Researchers from Epoch AI just showed that the open GLM model could have exploited known but still-unpatched Firefox vulnerabilities.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
原文版权归 The Decoder 及原作者所有，本站转载仅供阅读。如需删除，请联系 support@aisa.one。
Studio