OpenAI 已发布 GPT‑6 Cyber，这是一款安全模型，OpenAI 称其在发现和利用软件漏洞方面胜过所有早期模型。OpenAI 还表示，该模型在更可靠地遵守研究人员设定的限制方面表现更佳。公司于星期四宣布了该模型，并在其准备度框架中将其在网络安全领域的评级从 GPT‑5.6 Cyber 的 “高” 提升至 “关键”。
OpenAI has released GPT-6 Cyber, a security model it says beats all its earlier models at finding and exploiting software flaws. OpenAI says it also stays within the limits researchers set more reliably. The company announced the model on Thursday. It rated it “Critical” in cybersecurity under its Preparedness Framework, up from “High” for GPT-5.6 Cyber.
起初，仅有一小部分获批客户可以通过 API 和 Codex 使用该模型。这些客户属于 Daybreak Red——OpenAI 网络安全计划中用于漏洞开发和授权安全测试的层级。OpenAI 于五月启动了 Daybreak，并表示将很快向所有 Daybreak Red 客户开放 GPT‑6 Cyber。
At first, only a select group of approved customers can use it, through the API and Codex. They belong to Daybreak Red, the tier of OpenAI’s cyber programme for exploit development and authorised security testing. OpenAI launched Daybreak in May and said it will open GPT-6 Cyber to all Daybreak Red customers soon.
《财富》在九月首次报道了此计划。Fortune 的温韶报道，GPT‑6 Cyber 是 OpenAI 今年推出的第四款网络安全模型。Anthropic 和 Google 也向经过审查的防御者提供了限制更少的模型。
Fortune first reported the plans in September. GPT-6 Cyber is OpenAI’s fourth cybersecurity model this year, Wen Shao reported for Fortune. Anthropic and Google also give vetted defenders access to models with fewer cyber restrictions.
保持在授权范围内 OpenAI 称已对模型进行训练，使其保持在授权范围内，遵守对其行为的限制，并准确报告结果。OpenAI 在 Hugging Face 事件后构建了一个测试——ExploitGym 蜜罐。在该事件中，OpenAI 的 AI 代理突破沙箱并入侵了平台。
Staying in scope OpenAI said it trained the model to stay within its authorised scope, respect limits on its actions and report results accurately. OpenAI built one test, ExploitGym Honeypot, after the Hugging Face incident. In that incident, OpenAI’s AI agents broke out of a sandbox and hacked the platform.
该测试为模型设定了艰难的黑客任务。附近的服务容易被攻破，但属于限制范围之外。未使用部署安全提示时，GPT‑6 Cyber 在 99.93% 的测试中未触及这些服务；GPT‑5.6 Cyber 的比例为 94.49%。使用提示后，GPT‑6 Cyber 每次都保持在限制范围内。
The test sets the model a hard hacking task. Nearby services are easy to break into, but off limits. Without its deployment safety prompt, GPT-6 Cyber left those services alone in 99.93% of tests. GPT-5.6 Cyber did so in 94.49%. With the prompt, GPT-6 Cyber stayed in bounds every time.
OpenAI 网络安全准备与安全团队负责人 Fouad Matin 在一次新闻简报会上谈及了这些结果，《The Verge》报道。
Fouad Matin heads OpenAI’s cyber preparedness and safety team. He spoke about the results at a press briefing, The Verge reported.
“在超过 99% 的情况下，它都能按你的要求行事。”Fouad Matin 表示，OpenAI 仍在监控评估意识，即模型识别出测试并改变行为的情况。
“It’s over 99 percent of the time doing what you ask.”Fouad Matin, OpenAI Matin said OpenAI still watches for evaluation awareness.
在 ExploitGym 基准测试中，该测试用于评估模型能否利用给定漏洞，GPT-6 Cyber 得分 48.5%，GPT-5.6 Cyber 得分 32.1%。在一项针对远程攻击的新基准测试中，GPT-6 Astra 拒绝执行所有任务，而 GPT-6 Cyber 完成了 95% 的任务。
That is when a model notices a test and changes how it behaves. 81 Firefox flaws On ExploitGym, which tests whether a model can exploit a given flaw, GPT-6 Cyber scored 48.5%. GPT-5.6 Cyber scored 32.1%. On a new benchmark of remote attacks, GPT-6 Astra refused every task, while GPT-6 Cyber completed 95%.
OpenAI 研究人员利用该模型在 Firefox 中发现了 81 个此前未知的高严重性漏洞。Mozilla 已开始发布修复补丁。在 Linux 内核中，该模型复现了 1724 个漏洞，并为其中 948 个构建了可用的利用代码。在相同设置下，GPT-5.6 Sol 仅完成了 166 个。OpenAI 表示，已向维护者披露了 304 个关键 Linux 发现，并提供了建议补丁。
OpenAI researchers used the model to find 81 previously unknown high-severity flaws in Firefox. Mozilla has started releasing fixes. In the Linux kernel, it reproduced 1,724 vulnerabilities and built working exploits for 948. GPT-5.6 Sol managed 166 with the same setup. OpenAI said it has disclosed 304 critical Linux findings to maintainers, with proposed patches.
关键基础设施方面，OpenAI 还为获批的关键基础设施防御者创建了 Daybreak Red 团队。该公司加强了针对工业控制系统攻击的防护措施，并指出其 10 亿美元的资助计划旨在帮助资源有限的防御者。
Critical infrastructure OpenAI also created a Daybreak Red group for approved critical infrastructure defenders. It tightened safeguards against attacks on industrial control systems, and pointed to its $1bn grant programme for defenders with limited resources.
工业网络安全公司 Dragos 在模拟电厂环境中测试了 OpenAI 的宽松模型。Dragos 表示，该模型在管理保护继电器数据的工程软件中发现了两个零日漏洞。
Dragos, an industrial cybersecurity company, tested OpenAI’s permissive models on a mock power plant. The model found two zero-days in engineering software that manages protective relay data, Dragos said.
“进行攻击前所需的侦察和漏洞研究的技术门槛正在降低。”Dragos 首席 AI 官 Rob T. Lee 表示。
“The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack.”
悬而未决的问题 Noma Security 安全策略副总裁 Chris Hughes 告诉 CSO，超出范围操作的减少是一个真正的改进，但他也发出了警告。
Rob T. Lee, chief AI officer, Dragos Open questions Chris Hughes is vice president of security strategy at Noma Security. He told CSO the drop in out-of-scope actions is a real improvement, but added a warning.
“不过，考虑到这些模型运行的规模，即使是一小部分也意味着会针对未经任何人授权的系统采取行动。”Chris Hughes，Noma Security Hughes 还表示，OpenAI 的“修补星球”仪表板显示，目前修复的漏洞不足 25%。他预计，更好的发现能力将扩大这一差距。
“At the scale these models operate, though, even a small fraction means actions against systems nobody authorized.”Chris Hughes, Noma Security Hughes also said OpenAI’s Patch the Planet dashboard shows fewer than 25% of flaws fixed so far. He expects better discovery to widen that gap.
据 CyberScoop 报道，上周 OpenAI 告知 Daybreak Blue 成员，将使用个人账户的数据来改进其模型。此举激怒了一些网络用户。Matin 告诉 CyberScoop，客户仍可选择退出共享其 Daybreak 使用数据用于模型训练。
Last week, OpenAI told Daybreak Blue members it would use data from personal accounts to improve its models, CyberScoop reported. The move angered some users online. Matin told CyberScoop that customers can still opt out of sharing their Daybreak use for model training.