OpenAI 与 AWS 加速网络防御：Daybreak Red 和 Daybreak Blue 面向符合条件客户登陆 Amazon Bedrock
英文标题：Accelerate cyber defense with OpenAI and AWS: Daybreak Red & Daybreak Blue now available to eligible customers on Amazon Bedrock
OpenAI 的专业网络防御模型 Daybreak Red 和 Daybreak Blue 现已在 Amazon Bedrock 向符合条件的客户开放，分别提供 GPT-5.6 Cyber 和带防御安全护栏的 GPT-5.6 Sol。两款模型可用于漏洞发现、检测工程、事件响应、漏洞研究和修复方案开发，并通过身份验证、监控及访问控制处理网络安全任务的双重用途风险。AWS 表示，模型推理采用芯片级零运营人员访问（ZOA）、客户管理 KMS 密钥、IAM、CloudTrail 和 VPC 端点等机制，代码与漏洞数据不会用于模型训练。该服务目前仅在美国东部（弗吉尼亚北部）区域提供，客户需先通过 OpenAI 的 Trusted Access for Cyber 审核。
1.3 MB
101.1 KB
124.1 KB
345.0 KB
141.2 KB
346.5 KB
806.6 KB
8.2 KB
5.3 MB
219.1 KB
中文
网络防御人员从未拥有如此强大的能力，也从未如此迫切地需要这些能力。前沿模型如今能够对整个代码库进行推理，追溯漏洞的根本原因，并在数分钟内提出修复方案。对手同样可以获得这些能力。这正是漏洞披露到被利用之间的时间窗口不断缩短的原因。防御人员需要在庞大且陌生的代码库中跟上节奏，验证哪些发现真正重要，并交付经得起考验的补丁。他们还必须在自己能够控制和审计的环境中，妥善保管敏感代码和漏洞数据。挑战已不再是发现潜在问题，而是确认哪些问题真实存在、修复它们，并在时间窗口关闭前完成这些工作。
AWS 始终秉持从一开始就内建安全的原则。AI 也是如此：客户可以在 Amazon Bedrock 上运行从推理到完全自主 Agent 的各类工作负载，并使用他们在整个 AWS 中所依赖的同一套基础设施、控制措施和治理机制。如今，AWS 与 OpenAI 正将这一基础扩展至网络防御领域。OpenAI 的 Daybreak Red 和 Daybreak Blue 现已在 Amazon Bedrock 向符合条件的客户开放。Daybreak Red 提供针对网络安全专门训练的 GPT-5.6 Cyber 模型访问权限。Daybreak Blue 则提供 GPT-5.6 Sol 的访问权限，并配备针对防御性网络安全工作校准的安全护栏。两者均属于 OpenAI 的网络防御计划 Daybreak；该计划为防御人员提供受治理的前沿 AI 访问能力，包括 Agent 工具、应用红队测试，以及帮助防御人员将发现转化为经过测试的修复方案的服务。
“AWS 与 OpenAI 都相信，防御人员应当占据优势。这项合作将 OpenAI 的 Daybreak Red 和 Daybreak Blue 引入 Amazon Bedrock。AWS 安全团队如今正在使用这两款模型分析源代码、发现漏洞并开展红队研究。在 Bedrock 上，这些工作运行于 AWS 为所有其他关键工作负载采用的同一套基础设施控制之下。随着这些模型持续进步，防御人员在 Amazon Bedrock 上能够完成的工作也将随之不断拓展。”
— AWS Security 副总裁 John Sheehan
面向高级、获授权网络安全工作的专用模型
从发现漏洞到完成修复的路径很长：确认可利用性、追踪易受攻击组件如何可被访问、开发既能解决根本原因又不引入回归问题的修复方案、在真实条件下验证方案有效性，以及在披露窗口期内发布补丁。防御人员需要能够加速每一个环节的模型。但网络安全天然具有双重用途。无论意图为何，复现漏洞或逆向分析漏洞利用链的请求看起来完全相同，通用模型通常会通过拒绝请求来处理这种歧义。Daybreak Red 和 Daybreak Blue 则通过上下文来解决这一问题：谁在使用模型、工作在何处进行，以及哪些安全护栏在约束该访问。
对于大多数安全团队，Daybreak Blue 是合适的起点。它支持漏洞发现、检测工程和事件响应。Daybreak Red 则专为漏洞研究、漏洞利用复现和缓解措施开发等高级任务设计。对于这些任务，更低的拒绝阈值（即模型拒绝请求的临界点），配合更严格的身份验证、监控和访问控制，可提升调查的速度和深度。
OpenAI 表示，安全研究人员通过 Daybreak Red 使用 GPT-5.6 Cyber，发现了 Chrome 所使用的 JavaScript 引擎 V8 中两个此前未知的漏洞；这两个漏洞若被串联利用，可能导致内存损坏和堆沙箱逃逸。初始漏洞已被修复并以 CVE-2026-15903 发布，成为 2026 年 V8 CTF 中仅有的四个成功零日漏洞提交之一。
在您已信任的基础设施上运行
网络安全工作负载会向模型输入最敏感的数据，包括专有源代码、尚未修复的漏洞细节，以及来自生产系统的实时遥测数据。在将这些数据交由模型处理前，您需要明确知道数据会流向何处，以及谁可以查看它们。
两款模型均运行在为高性能、安全性和可靠性打造的 Amazon Bedrock 下一代推理引擎上。芯片级强制执行零运营人员访问（ZOA），因此即使 AWS 运营人员也无法在推理过程中访问您的提示词和生成结果。所有数据在传输中和静态存储时均使用客户管理的 AWS Key Management Service（AWS KMS）密钥加密。访问权限由您的 AWS Identity and Access Management（IAM）策略控制、在 AWS CloudTrail 中记录，并通过虚拟私有云（VPC）端点路由。您还可以在组织层面设置数据边界策略，防止数据跨账户和网络边界外泄。
您的推理数据不会被用于模型训练，两款模型也均不要求您选择与 OpenAI 共享数据。对于自动化滥用检测，由分类器标记的流量会由 AWS 最长保留 30 天，并以程序化方式处理。客户可通过其 AWS 客户团队申请零数据保留。详情请参阅数据保留说明。
开始使用
Daybreak Red：GPT-5.6 Cyber，以及 Daybreak Blue：GPT-5.6 Sol，现已在以下 AWS 区域的 Amazon Bedrock 向符合条件的客户开放：美国东部（弗吉尼亚北部）。访问这些模型需要先加入 OpenAI 的 Trusted Access for Cyber 计划。若要加入，请联系 OpenAI，或联系您的 AWS 客户团队以获取资格相关指导。获批后，请与您的客户团队合作申请 AWS 访问权限。更多信息请参阅文档。
如需探索 Amazon Bedrock 上完整的 GPT-5.6 系列，请参阅在 Amazon Bedrock 上开始使用 OpenAI GPT-5.6 Sol、Terra 和 Luna。
想了解 Amazon Bedrock 如何支持您的团队？请联系我们，开启交流。
关于作者
Tanvi Girinath
Tanvi 是 Amazon Web Services（AWS）Amazon Bedrock 的产品营销经理，负责帮助客户借助 Amazon Bedrock 采用并规模化部署 AI 应用和 Agent。
Saurabh Trikande
Saurabh 是 Amazon Web Services（AWS）Amazon Bedrock 的高级产品经理。他负责推动领先提供商的前沿模型推理能力，使其能够以高性能、安全且高成本效益的方式服务于任何规模的客户。
Chris Dickens
Chris 是 OpenAI 的产品团队成员，专注于 OpenAI API。他的工作包括与 AWS 围绕 Amazon Bedrock 开展合作，使开发者能够广泛使用 OpenAI 的前沿模型。
English / 原文
Cyber defenders have never had more capability at their fingertips, and they have never needed it more. Frontier models can now reason across an entire code base, trace a vulnerability to its root cause, and propose a fix in minutes. Those same capabilities are available to adversaries. This is why the window between a vulnerability being disclosed and being exploited keeps shrinking. Defenders are expected to keep pace across sprawling, unfamiliar code bases, validate which findings matter, and ship patches that hold. They must do all of this while keeping sensitive code and vulnerability data inside an environment they control and can audit. The challenge is no longer surfacing potential issues but confirming which ones are real, fixing them, and doing it before the window closes.
AWS has always operated on the principle that security is built in from the start. The same is true for AI: customers run workloads on Amazon Bedrock, from inference to fully autonomous agents, under the same infrastructure, controls, and governance they rely on across AWS. Today, AWS and OpenAI are extending that foundation to cyber defense. Daybreak Red and Daybreak Blue from OpenAI are now available on Amazon Bedrock to eligible customers. Daybreak Red provides access to GPT-5.6 Cyber, a purpose-trained cybersecurity model. Daybreak Blue provides access to GPT-5.6 Sol with safeguards calibrated for defensive cybersecurity work. Both are part of Daybreak, the cyber defense initiative from OpenAI that gives defenders governed access to frontier AI, including agentic tooling, application red teaming, and services that help defenders move from findings to tested fixes.
“AWS and OpenAI share a belief that defenders should have the advantage. This partnership brings Daybreak Red and Daybreak Blue from OpenAI to Amazon Bedrock. AWS security teams are using both models today to analyze source code, discover vulnerabilities, and conduct red-team research. On Bedrock, that work runs under the same infrastructure controls AWS applies to every other critical workload. As these models continue to advance, so will what defenders can do with them on Amazon Bedrock.”
— John Sheehan, Vice President, AWS Security
Specialized models for advanced, authorized cyber security work
The path from vulnerability discovery to remediation is long: confirming exploitability, tracing how the vulnerable component is reachable, developing a fix that addresses root cause without introducing regressions, validating it holds under realistic conditions, and shipping a patch within the disclosure window. Defenders need models that can accelerate every step. But cybersecurity is inherently dual use. A request to reproduce a vulnerability or reverse engineer an exploit chain looks identical regardless of intent, and general-purpose models resolve that ambiguity by declining the request. Daybreak Red and Daybreak Blue resolve it through context: who is using the model, where the work occurs, and what safeguards govern that access.
Daybreak Blue is the right starting point for most security teams. It supports vulnerability discovery, detection engineering, and incident response. Daybreak Red is designed for advanced tasks like vulnerability research, exploit reproduction, and mitigation development. For these tasks, a lower refusal threshold (the point at which the model declines a request) matched by stronger identity verification, monitoring, and access controls improves the speed and depth of an investigation.
According to OpenAI, security researchers used GPT-5.6 Cyber through Daybreak Red to identify two previously unknown vulnerabilities in V8, the JavaScript engine used by Chrome, which when chained together could enable memory corruption and a heap sandbox escape. The initial vulnerability was fixed and released as CVE-2026-15903, one of only four successful zero-day entries to V8 CTF in 2026.
Run it on the foundation you already trust
A cyber security workload feeds the model your most sensitive inputs: proprietary source code, unpatched vulnerability details, and live telemetry from production systems. Before any of that runs through a model, you need certainty about where it goes and who can see it.
Both models run on the Amazon Bedrock next-generation inference engine built for high performance, security, and reliability. Zero-operator access (ZOA) is enforced at the chip, so even AWS operators cannot access your prompts and completions during inference. Everything is encrypted in transit and at rest with customer-managed AWS Key Management Service (AWS KMS) keys. Access is governed by your AWS Identity and Access Management (IAM) policies, logged in AWS CloudTrail, and routed through virtual private cloud (VPC) endpoints. You can set data perimeter policies at the organization level to prevent exfiltration across account and network boundaries.
Your inference data is not used for model training, and neither model requires you to opt into sharing your data with OpenAI. For automated abuse detection, classifier-flagged traffic is retained by AWS for up to 30 days and processed programmatically. Customers may request zero data retention through their AWS account team. See data retention for details.
Get started
Daybreak Red: GPT-5.6 Cyber and Daybreak Blue: GPT-5.6 Sol are now available to eligible customers on Amazon Bedrock in the following AWS Region: US East (N. Virginia). Access to the models requires enrollment in Trusted Access for Cyber from OpenAI. To enroll, contact OpenAI or reach out to your AWS account team for guidance on eligibility. Once approved, work with your account team to request access on AWS. To learn more, see the documentation.
To explore the full GPT-5.6 family on Amazon Bedrock, see Get started with OpenAI GPT-5.6 Sol, Terra, and Luna on Amazon Bedrock.
Interested in how Amazon Bedrock can support your team?Connect with us to start the conversation.
About the authors
Tanvi Girinath
Tanvi is a Product Marketing Manager for Amazon Bedrock at Amazon Web Services (AWS), where she helps customers adopt and scale AI applications and agents with Amazon Bedrock.
Saurabh Trikande
Saurabh is Senior Product Manager for Amazon Bedrock at Amazon Web Services (AWS). He leads efforts to make inference with frontier models from leading providers performant, secure, and cost-efficient for customers at any scale.
Chris Dickens
Chris is a Member of Product Staff at OpenAI focused on the OpenAI APIs. His work includes collaboration with AWS on Amazon Bedrock to make OpenAI’s frontier models widely accessible to developers.