OpenAI 安全模型发现并利用 948 个 Linux 内核提权漏洞
OpenAI 用 AI 扫 Linux 内核，找到 1724 个真漏洞还写出了 948 个提权 exploit，一个漏洞成本才 1279 美元，搞安全的都该看看。
OpenAI 的 Scout agents 将 Linux 内核 4000 多万行代码筛选出 280 万行最值得审计的部分，即约 7%。这批 AI agents 标记出 2308 个疑似漏洞，在测试机中确认 1724 个，并针对其中 948 个构建了可用的 root 提权 exploit 和修复补丁，结果经人工复核后提交给 Linux 维护者。整个流程消耗 1,213,430 美元 API 额度，其中 775,929 美元用于构建 exploit。按此计算，用 AI 找到并武器化一个 Linux 内核漏洞的成本约为 1,279 美元。
OpenAI claims their new security model patched 948 Linux privilege escalation vulnerabilities.
Scout agents cut the kernel’s 40M-plus lines to 2.8M worth auditing. i.e. AI agents picked the most security-relevant 7% of the Linux kernel (2.8M lines), flagged 2,308 suspected bugs, confirmed 1,724 in test machines, and built working root-takeover exploits and fixes for 948. Humans reviewed the results before sending them to Linux maintainers.
The whole run cost $1,213,430 in API credits, $775,929 of it on building exploits, though OpenAI ran its own model, so that’s likely not cash actually spent.
So all this means that finding and weaponizing a Linux kernel bug now costs about $1,279 with AI, so any unpatched Linux machine where an attacker gets even a basic user account should be treated as fully taken over.
