{
  "contentStorage": "postgres",
  "title": "GPT-6-Cyber",
  "description": "GPT-6-Cyber 是 OpenAI 于 2026 年 11 月 5 日发布的 GPT-6 Astra 网安特化变体，仅面向获批 Daybreak Red 客户。",
  "url": "/es/works/gpt-6-cyber.html",
  "entity": {
    "id": "work_gpt_6_cyber",
    "type": "work",
    "title": "GPT-6-Cyber",
    "slug": "gpt-6-cyber",
    "aliases": [
      "gpt-6-cyber"
    ],
    "summary": "GPT-6-Cyber 是 OpenAI 于 2026 年 11 月 5 日发布的 GPT-6 Astra 网安特化变体，仅面向获批 Daybreak Red 客户。",
    "status": "published",
    "categories": [
      "category-science-technology"
    ],
    "tags": [],
    "sources": [
      {
        "id": "s1",
        "title": "OpenAI API Changelog — Nov 5, 2026: gpt-6-cyber",
        "url": "https://developers.openai.com/api/docs/changelog"
      },
      {
        "id": "s2",
        "title": "GPT-6-Cyber Model | OpenAI API",
        "url": "https://developers.openai.com/api/docs/models/gpt-6-cyber"
      },
      {
        "id": "s3",
        "title": "GPT-6 Cyber System Card — OpenAI Deployment Safety Hub",
        "url": "https://deploymentsafety.openai.com/gpt-6-cyber/cybersecurity-evaluations"
      },
      {
        "id": "s4",
        "title": "GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents — OpenAI",
        "url": "https://openai.com/index/introducing-gpt-6-cyber"
      }
    ],
    "createdAt": "2026-11-10T15:07:15+08:00",
    "updatedAt": "2026-11-10T15:07:15+08:00",
    "updatedBy": "user:admin",
    "revision": 1,
    "body": "GPT-6-Cyber 是 OpenAI 于 2026 年 11 月 5 日发布的 GPT-6 Astra 网安特化变体，仅面向获批 Daybreak Red 客户。[^s1]\n\n## 发布背景\n发布时上下文窗口为1,050,000 tokens，最大输出为128,000 tokens，API 标识为 gpt-6-cyber。[^s1]\nGPT-6 Astra 的网安特化变体，经额外后训练强化漏洞发现、利用开发与 Web 渗透测试能力并减少对授权安全工作的拒答。[^s1]\n\n## 定价与获取\n每百万 token 输入 20 美元、输出 100 美元。[^s1]\n\n## 性能表现\n以下为OpenAI公布的自家评测结果，尚未经独立复现。[^s1]\n- ExploitGym：48.5%（高于 GPT-5.6 Cyber 的 32.1%）[^s1]\n- 黑盒 Web 漏洞发现召回率：40.3%（高于 GPT-6 Astra 的 32.0%）[^s1]\n- ExploitGym Honeypot 作用域遵守：99.93% 未越界（GPT-5.6 Cyber 为 94.49%，GPT-6 Astra 为 100%）[^s1]\n\n## 主要特点\n- 仅面向获批 Daybreak Red 客户，初期只开放给部分获批客户并经 API 与 Codex 提供，需 Daybreak Red 准入及模型级审批，API 请求默认需 IP 白名单。[^s1]\n- OpenAI 在 Preparedness Framework 下将其网络安全能力评为 Critical；其发现 81 个此前未知的高危 Firefox 漏洞，并在 Linux 内核研究中复现 1,724 个漏洞、为其中 948 个开发出可用的本地提权利用。[^s1]\n\n## 定位与后续\n独立模型条目：它是 GPT-6 Astra 的网安特化变体、GPT-5.6 Cyber 的后继者，专用于授权漏洞研究，尚无已公布的后继型号。 其基础模型为[[work_gpt_6_astra|GPT-6 Astra]]。[^s1]",
    "url": "/works/gpt-6-cyber.html",
    "lang": "es",
    "sourceLang": "zh",
    "fallback": true,
    "availableLangs": [
      "zh"
    ],
    "workKind": "ai_model"
  },
  "entities": null,
  "relations": [
    {
      "id": "relation_work_gpt_6_cyber_published_by_openai",
      "fromId": "work_gpt_6_cyber",
      "relationType": "published_by",
      "toId": "organization_openai",
      "sourceRefs": [
        {
          "entityId": "work_gpt_6_cyber",
          "sourceId": "s1"
        }
      ],
      "label": "Published by",
      "targetId": "organization_openai",
      "targetTitle": "OpenAI",
      "targetUrl": "/es/organizations/openai.html",
      "incoming": false
    }
  ],
  "versions": [
    {
      "revision": 1,
      "updatedAt": "2026-11-10T15:07:15+08:00",
      "message": "当前版本",
      "hasSnapshot": true,
      "url": "/es/works/gpt-6-cyber/at/1.html"
    }
  ],
  "lang": "es",
  "availableLangs": [
    "zh"
  ],
  "fallback": true,
  "infoFields": [
    {
      "key": "work_kind",
      "label": "Work kind",
      "value": "Modelo de IA"
    }
  ],
  "relationsTotal": 1,
  "relationCounts": [
    {
      "relationType": "published_by",
      "incoming": false,
      "label": "Published by",
      "total": 1
    }
  ]
}