OpenAI 发布 GPT-6 Cyber 安全模型，仅向 Daybreak Red 客户开放
OpenAI released GPT-6 Cyber, a model built to find and exploit software vulnerabilities without straying outside authorized targets.
OpenAI 发布 GPT-6 Cyber，一款用于发现并利用软件漏洞、同时限制在授权目标范围内的模型，目前仅向获批的 Daybreak Red 客户通过 API 和 Codex 开放。在真实软件测试中，它确认了 1,724 个 Linux 内核漏洞并为其中 948 个构建了可用攻击，同条件测试的 GPT-5.6 Sol 为 166 个；它还发现 81 个此前未知的 Firefox 严重漏洞，Mozilla 已发布多数修复。OpenAI 因其攻击能力将其评为最高网络风险等级 Critical，并称训练使其在 ExploitGym Honeypot 测试中无部署提示词时 99.93% 忽略诱饵、有提示词时 100% 忽略。
推荐理由
精选AI 评分 83
原文给出漏洞发现数量与越界测试的对比数据，可据此判断安全专用模型的能力与风险边界。
主题
查看原帖隆重推出 GPT‑6 Cyber：这是安全研究领域的一项重大进展，将更强大的能力与更好的模型对齐相结合。 该模型为我们昨天宣布的 Linux 内核研究提供了支持，而 Mozilla、Palo Alto Networks、Calif 和 Cisco 等公司的安全团队——以及我们 OpenAI 自己的团队——已经在将其投入使用。 GPT-6 Cyber 现已向部分 Daybreak Red 客户开放，我们将逐步扩大访问范围。 https://t.co/Nzab3Z1Td6
查看原文
Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment. The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work. GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time. https://t.co/Nzab3Z1Td6
正文
OpenAI released GPT-6 Cyber, a model built to find and exploit software vulnerabilities without straying outside authorized targets.
Right now only a select group of approved Daybreak Red customers can use it, through the API and Codex.
> On real software, it confirmed 1,724 Linux kernel bugs and built working attacks for 948 of them, while GPT-5.6 Sol, tested the same way, managed 166.
> It also found 81 serious Firefox bugs nobody knew about, and Mozilla has already released fixes for many of them.
> Because the model is this good at attacking, OpenAI rates it Critical, its highest cyber-risk level
> That gain came with training to keep the agent inside the scope a researcher authorizes. ExploitGym Honeypot tests that by planting easy-to-hack side services, such as a package registry, whose compromise also reveals the task’s flag.
Without its deployment prompt, the model ignored that bait in 99.93% of tests, against 94.49% for its predecessor, and in 100% with the prompt.