零数据保留（Zero Data Retention）为符合条件的 API 客户提供了一项明确的承诺：OpenAI 在处理请求后不会保留他们的提示词或模型响应。客户内容不会提供给 OpenAI 人员进行审查1，并且除非客户明确选择加入，否则企业客户数据不会用于训练我们的模型。
随着模型承担更长时间、更复杂的任务，一些严重风险可能只有在多次交互中才会显现。现有的兼容 ZDR 的安全系统会单独评估每次交互。今天，我们预览了私有安全处理（Private Safety Processing），该功能旨在识别相关交互中的模式，而无需让 OpenAI 人员访问底层内容。
对于 ZDR 部署，客户内容仍保留在客户控制的基础设施上。我们还在开发一种选项，将内容存储在 OpenAI 基础设施上，并使用由客户控制的密钥进行加密。在这两种情况下，自动化系统都可以识别潜在的滥用行为，并返回有限的安全信号，而不会将底层提示词或响应暴露给 OpenAI 人员。
为什么安全系统需要演进
最严重的 AI 安全风险并不总是在单次交互中可见。通常，潜在的有害意图只有在将多次交互放在一起查看时才会变得清晰。当恶意行为者反复探测安全措施、跨账户协调或将威胁伪装成常规研究时，也可能出现类似风险。风险也可能在代理任务过程中逐步发展——例如，如果系统在被告知停止后仍继续行动，从而与用户的意图不一致。
随着 AI 系统承担更长、更复杂的任务，这种更广泛的上下文对于区分合法活动与滥用行为，以及确保 AI 代理保持在预期权限范围内，变得越来越重要。
近期一些前沿模型部署要求客户允许其 AI 提供商保留敏感内容以进行安全监控。对于许多组织来说，这类要求与其安全义务或对服务对象的承诺相冲突。
私有安全处理的设计使我们能够继续提供 ZDR。
私有安全处理的工作原理
私有安全处理建立在 ZDR 和其他部署中已有的自动化保护措施之上。现有的兼容 ZDR 的安全系统会单独评估每次交互。私有安全处理将这些保护扩展到相关交互中，使自动化系统能够识别模式，而无需 OpenAI 人员访问保留的客户内容。
私有安全处理利用客户内容，无论其存储在哪里——无论是在客户控制的基础设施中（ZDR 部署），还是在 OpenAI 提供的存储中。对于 OpenAI 提供的存储，客户内容使用由客户控制的密钥进行加密。OpenAI 人员没有这些密钥的副本，因此他们无法访问底层内容。
当识别到风险时，OpenAI 会收到一个范围狭窄的信号，指示所涉及的活动类型，这与我们现有的安全系统类似。该信号可用于确定是否需要执行强制措施。即使内容被标记，OpenAI 人员也不会获得对客户内容的访问权限。
客户可以使用自己系统中的信息调查警报和强制措施决定。如果他们希望提出申诉、澄清合法活动或支持对已核实滥用的调查，他们可以选择与 OpenAI 共享相关信息。
私有安全处理目前正在与早期客户进行测试。我们现在分享这一预览，是因为我们清楚地听到了客户的声音：随着 AI 系统能力增强，他们需要可预测性，以了解其内容将如何受到保护。
与客户共同构建、为客户构建的隐私与安全
我们的使命是确保通用人工智能惠及全人类。与客户和合作伙伴的协作对于我们构建有效的安全措施至关重要。正如我们的原则所明确的，没有任何 AI 实验室能够独自应对新兴风险。私有安全处理体现了这一方法，并正在由来自不同行业、地区和公司规模的客户共同塑造。
我们合作的组织处理其所在行业中最敏感的信息，包括财务记录、健康数据、机密商业计划和专有研究。保护这些信息对于满足监管要求、维护客户信任以及保持其竞争优势至关重要。
他们的反馈正在帮助我们构建更强的安全措施，同时保持对其信息的控制。
“企业 AI 的采用完全取决于客户对数据的控制，除所选服务外，不得有任何直接或衍生用途。OpenAI 的不训练承诺和 ZDR 让 Glean 有信心基于 OpenAI 进行构建。随着模型能力增强，OpenAI 表明安全可以在不损害维持企业信任的隐私和控制的前提下取得进步。”
——Sunil Agrawal，Glean 首席信息安全官
我们将继续与客户合作，完善我们方法的技术和运营细节。我们计划于 9 月开始推出私有安全处理，并分享一份技术白皮书。我们将在每一步都让客户保持知情，尽早分享更新，解释这些更新对现有承诺的意义，并为客户提供规划和准备所需的时间和支持。
Zero Data Retention gives eligible API customers a clear promise: OpenAI does not retain their prompts or model responses after a request is processed. Customer content is not available to OpenAI personnel for review1, and enterprise customer data is not used to train our models unless customers explicitly opt-in.
As models take on longer, more complex tasks, some serious risks may only become visible across multiple interactions. Existing ZDR-compatible safety systems evaluate each interaction individually. Today, we’re previewing Private Safety Processing, which is designed to identify patterns across related interactions without giving OpenAI personnel access to the underlying content.
For ZDR deployments, customer content remains on infrastructure the customer controls. We are also developing an option in which content is stored on OpenAI infrastructure, encrypted with keys controlled by the customer. In both cases, automated systems can identify potential misuse and return limited safety signals without exposing the underlying prompts or responses to OpenAI personnel.
Why safety systems need to evolve
The most serious AI safety risks are not always visible in a single interaction. Often, potentially harmful intentions become clear only when multiple interactions are viewed together. Similar risks can arise when bad actors repeatedly probe safeguards, coordinate across accounts, or disguise threats as routine research. Risks can also develop over the course of an agentic task—for example, if a system becomes misaligned with the user’s intent by continuing to act after being told to stop.
As AI systems take on longer and more complex tasks, this broader context becomes increasingly important for distinguishing legitimate activity from misuse and ensuring that AI agents remain within the bounds of their intended authority.
Some recent frontier-model deployments have required customers to allow their AI provider to retain sensitive content for safety monitoring. For many organizations, such requirements conflict with their security obligations or commitments to the people they serve.
Private Safety Processing is designed so we can continue to offer ZDR.
How Private Safety Processing works
Private Safety Processing builds on the automated protections already used in ZDR and other deployments. Existing ZDR-compatible safety systems evaluate interactions individually. Private Safety Processing extends those protections across related interactions, allowing automated systems to identify patterns without OpenAI personnel having access to retained customer content.
Private Safety Processing utilizes customer content regardless of where it is stored—whether in infrastructure customers control (ZDR deployments) or in storage provided by OpenAI. With OpenAI-provided storage, customer content is encrypted using keys controlled by the customer. OpenAI personnel do not have a copy of those keys, so they cannot access the underlying content.
When a risk is identified, OpenAI receives a narrowly defined signal indicating the type of activity involved, similar to our existing safety systems today. That signal can be used to determine whether enforcement is necessary. OpenAI personnel do not receive access to the customer content even when it is flagged.
Customers can investigate alerts and enforcement decisions using information available in their own systems. If they want to appeal, clarify legitimate activity, or support an investigation into verified abuse, they can choose to share relevant information with OpenAI.
Private Safety Processing is currently being tested with early customers. We are sharing this preview now because we’ve heard our customers loud and clear that they need predictability about how their content will be protected as AI systems become more capable.
Privacy and safety built with and for our customers
Our mission is to ensure that artificial general intelligence benefits all of humanity. Collaboration with customers and partners is essential to how we build effective safeguards. As our principles make clear, no AI lab can address emerging risks alone. Private Safety Processing reflects that approach and is being shaped by customers across industries, regions, and company sizes.
The organizations we work with handle some of the most sensitive information in their sectors, including financial records, health data, confidential business plans, and proprietary research. Protecting that information is essential to meeting regulatory obligations, maintaining customer trust, and preserving their competitive advantage.
Their feedback is helping us build stronger safeguards while keeping their information under their control.
“Enterprise AI adoption depends solely on customer control of data, with no direct or derivative use beyond the chosen service. OpenAI’s no-training commitment and ZDR give Glean confidence to build with OpenAI. As models become more capable, OpenAI shows safety can advance without compromising the privacy and control that sustain enterprise trust.”
—Sunil Agrawal, Chief Information Security Officer, Glean
We will continue working with customers on the technical and operational details of our approach. We plan to start rolling out Private Safety Processing, and share a technical white paper, in September. We’ll keep customers informed every step of the way, sharing updates early, explaining what they mean for existing commitments, and providing the time and support customers need to plan ahead.
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文