← 返回选刊

OpenAI / 文章

四家独立机构将进入OpenAI内部,查看检查点、代码、训练数据和监控系统

OpenAI在2026年11月16日宣布,已与Apollo Research、METR、Transluce和Goodfire签约,启动第一阶段嵌入式评估。目前公开的是访问范围和合作规程,评估结论还没有发布。

乔木改写

OpenAI在2026年11月16日宣布,已经和四家独立评估机构签约,启动第一阶段嵌入式评估。这四家是Apollo Research、METR、Transluce和Goodfire。

官方把嵌入式评估写成:让独立专家在较长一段时间里,深入接触实验室内部的模型和系统,评估模型对齐情况,以及防护措施是否有效,尤其是和安全论证相关的部分。安全论证的原文是safety cases。

评估方能拿到哪些内部资源,要看每次评估需要什么。官方列出的范围,包括模型检查点、源代码、训练和评估数据、监控系统、技术文档,以及和负责这些工作的研究员、工程师直接讨论。这些资源被写成和OpenAI内部研究人员使用的部分材料相同。官方的理由是,评估方需要证据和上下文,才能独立做评测,并发现他们自己可能没预料到的失效。

第一阶段的重点,是模型在训练过程中和内部使用中的对齐与防护,用来补充OpenAI的安全论证。文章把这项工作追溯到两处已有安排:一份关于第三方评估优先事项和原则的说明,以及多年来和其他评估方的合作,其中包括METR最近一次对内部监控系统、并有深度内部访问的试点,还有METR对Hugging Face事件的调查。

官方还写,这些合作也支撑他们在白宫超级智能协议下的承诺。协议的原文是White House Accord on Super Intelligence,页面没有给出单独链接。承诺的内容,按官方表述,是同意对安全控制做独立评估。他们希望这类合作有助于形成更强的、可共享的独立评估标准和公开透明度。

目前公开的是签约内容和评估范围,评估结论还没有发布。

Apollo要分两阶段核验四个关于隐蔽对抗的主张

Apollo Research的试点,要评估模型会不会隐蔽地违背开发者意图,以及现有防护能不能发现这种行为。依据来自训练、安全测试和内部使用。官方原文在这里有一句不完整的英语句子,中文按随后列出的四项主张来写。

Apollo要评估的四项主张是:训练没有激励颠覆行为;强评估和红队测试也诱发不出颠覆行为;模型在内部使用中没有尝试过颠覆;如果出现颠覆,其推理过程应能被发现。评估分两个阶段进行。

第一阶段将提供的访问包括:部分已公开和未发布的模型及训练检查点,相关训练数据和评估结果,训练干预记录,监控系统及被标记的行为,能提供上下文的研究人员,以及独立测试所需的工具和算力。

Goodfire的主要读者是OpenAI内部研究员

Goodfire会派可解释性研究员进驻,并用Goodfire自己的平台分析模型内部。评估的主要目标是训练过程监控、调试和对齐。官方写,这些评估的首要读者是OpenAI内部研究员,他们会用这些发现去做安全改进。Goodfire也会在合作推进过程中公布学习和发现。

METR先看最大的一轮强化学习,更大范围的监控评估还没定范围

OpenAI打算和METR做两项评估,分别针对对齐和监控。

对齐评估的第一阶段,METR要在迄今最大的一轮强化学习中,寻找错误对齐的行为,包括现有测试和监控可能漏掉的失效。在选定的训练轮次、已商定的访问、资源和安全安排之内,METR可以独立设定评估优先级,并反复做检索、分析和诱发实验。访问范围包括抽样相关检查点、训练和评估展开过程,以及这一轮的错误对齐监控记录。官方在脚注里写,这个范围比METR在对齐评估方案里设想的广泛排查更窄。

监控评估方面,双方最近已经做过一次监控压力测试。脚注补充,那是一次为期3周的监控红队演练,METR当时获得了代码库、监控器、相关轨迹,以及和OpenAI研究同事的讨论,目的是更好地理解这些监控器。这次演练的公开说明在METR于2026年10月22日发布的博文里。

官方写,双方已初步同意再做一次范围打算更广的监控评估,但还要等到双方就评估范围达成一致。这份监控评估方案的链接,指向METR在2026年10月14日发布的控制评估提案。

Transluce先看训练后对齐评估的质量和覆盖,后续评估还只是意向

Transluce的初始评估,针对OpenAI训练后对齐评估的质量和覆盖面。初始阶段将提供部分已公开和未发布的模型,训练后对齐评估及其相关结果,支持性的模型行为数据和安全论证文档,能提供上下文的研究人员,以及独立测试所需的工具和算力。

OpenAI和Transluce还共同表示,打算在2027年初再做一次约八周的后续评估,访问更多模型,更深入地查看对齐评估,并了解评估结果如何影响模型发布决定。这是双方的意向,还不是已经开始的评估。

敏感信息要在OpenAI办公室内、用公司设备访问

官方说,这些规程是和评估方一起制订的,要在深度访问与安全、保密、监督,以及评估的独立性和完整性之间做平衡。他们公开这些安排,是为了说明合作方式,也想推动尚未标准化的嵌入式评估规范讨论。官方预期这些规程还会随经验调整,并预先提到早期可能遇到技术访问障碍、空间限制、安全规程变化和其他意外问题。

每次合作先商定范围,以及做严格评估所需的访问、资源和安全安排。有的评估是在检验特定安全主张,有的则是在既定系统或训练轮次里寻找未曾预料的失效模式。两种情况下,评估方都需要能跟着证据走,调整方法,并在发现需要时提议进一步评估。如果直接访问可能暴露最敏感的知识产权,例如模型权重或架构,官方会考虑替代办法,包括让评估方和OpenAI研究人员一起跑分析。

需要接触敏感保密信息的评估方,要在OpenAI办公室现场工作,并使用OpenAI提供的设备。官方把主要目的写成保护所访问信息的安全,同时现场工作也便于和专门对接的员工沟通,获得访问协助和上下文。嵌入式评估方在做实验和评测时,至少要遵守与OpenAI员工相同的安全标准。脚注补充,某些实验条件下会采用更严的标准,以便做更仔细的安全审查。

完整未删减报告会交给安全与安保委员会,员工谈话默认不公开

OpenAI基金会董事会下有一个独立的监督委员会,名为安全与安保委员会,英文是Safety and Security Committee,简称SSC。官方写,SSC可以在评估前、评估中和评估结束时履行监督职责,并会收到未删减的报告。评估方和SSC在评估期间可以互相交谈,包括私下交谈。

OpenAI与任一评估方的分歧,会先交给Safety Advisory Group和OpenAI高管,并在规定时限内处理;一旦升级,SSC会立刻得到通知。如果仍无法达成一致,事项将提交SSC裁决。

员工谈话被用来帮助评估方理解内部信息和背景。OpenAI会授权并指导可以和评估方交谈的员工,评估方也可以要求增加谈话对象。为了让谈话更坦率,并保护员工隐私,这些讨论是不公开记录的:它们可以影响评估,但不能被发表,也不能归于个别员工。

另有一条独立的证明路径。评估方可以要求OpenAI出具经过证明的书面说明,用来核实、纠正或澄清他们从员工谈话里了解到的内容。这些说明不归于个人,也不用来描述个人观点。评估方可以公布OpenAI同意证明的说明,也可以披露OpenAI拒绝证明某条说明。

报酬、冷却期和公开前的删减权,官方都写进了限制条件

利益冲突,包括私人关系、财务利益和其他问题,官方写成需要清楚、透明地管理。冷却期也需要仔细设计,好让评估方能在不同实验室之间负责任地工作,同时谨慎处理竞争信息。官方说,他们正在和评估方一起制订冲突和冷却期的明确流程,并允许在缓解措施上保留弹性,在要求回避的决定上保持透明。

脚注披露了报酬安排。Apollo Research和Goodfire接受工作报酬,但报酬不取决于评估结果。METR和Transluce不接受评估报酬,但可以接受评估中使用的API额度或算力。官方写,他们不认为接受报酬就会让评估方失去独立性,但认为有必要披露。

冷却期的脚注写,参与嵌入式评估的第三方机构人员,在再嵌入另一家AI开发者之前,需要遵守一段短暂的冷却期,以降低无意泄露保密信息的风险。官方没有给出这段冷却期的具体时长。

评估方在发布评估时保持独立,但OpenAI保留删减保密信息的权利,并写明这与以往第三方合作一致。官方也可能要求负责披露的延迟,以便有时间处理可能带来安保或安全风险的发现,并修补仍未解决的问题。

更详细的保密报告会交给监督机构。面向公众的报告,则被写成帮助人们理解发现及其限度。

第一阶段之后,官方还想把嵌入式评估做成更长期的安排

官方希望第一阶段评估,能导向更长期、持续的嵌入式合作,持续提供第三方安全信号,并对他们最吃重的安全主张和安全论证提出改进建议。他们还会继续与多家独立第三方合作,理由是要形成方法多样的评估生态。有的调查需要在现场、对着敏感系统做持续工作,有的主要评估已部署或即将部署的模型,或审阅公开文档。

接下来几个月,他们计划把外部评估扩展到更广的安全问题。官方还在探索未来的嵌入式评估,可能检查他们对安全框架的遵守情况,以及组织文化和流程。这些是正在探索的方向,不是已经签约的本轮范围。

如果要跟踪后续结果,可以直接看各评估方自己将发布的公开报告,同时也要记住这次安排里的限制:现场和公司设备、删减权、负责披露延迟、员工谈话默认不公开,以及部分评估范围仍待双方确认。OpenAI在同日发布的公告线程是:https://x.com/OpenAI/status/2122359930628063248

原文

Our first phase of embedded assessments

Our first phase of embedded assessments

November 16, 2026

We have signed agreements with four independent assessors—Apollo Research, METR, Transluce, and Goodfire—for our first phase of embedded assessments. As AI capabilities advance rapidly and the potential risks grow, public reporting from independent assessments can give people a clearer view of how labs are evaluating and managing those risks.

Embedded assessments give independent experts deep access to a lab’s internal models and systems to assess model alignment and the effectiveness of safeguards over an extended period of time, especially where relevant to our safety cases. Based on what each evaluation needs, assessors receive access to some of the same internal resources our researchers use, including model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the researchers and engineers responsible for the work. This access gives assessors the evidence and context to pursue independent evaluations and uncover failures we may not have anticipated.

Our initial assessments focus on alignment and safeguards for models during training and internal use, complementing our safety cases. They build on our priorities and principles for third party assessments and years of work with third party evaluators, including a recent pilot assessment of our internal monitoring systems with deep internal access and an investigation of the Hugging Face incident. We’ll use what we learn to deepen our work with independent experts in the coming months. These efforts also support our commitments under the White House Accord on Super Intelligence, where we agreed to independent assessments of our safety controls. These collaborations can contribute to stronger shared standards for independent assessment and public transparency across the industry.

Our pilot embedded assessors

Apollo Research

In this pilot assessment, Apollo Research will assess whether models could covertly work against their developers’ intentions and whether our safeguards would detect such behavior. Drawing on evidence from training, safety testing, and internal use. Apollo will assess four claims in two stages:

  1. Training did not incentivize subversion.
  2. Strong evaluations and red teaming do not elicit subversion.
  3. Models have not attempted subversion during internal use.
  4. Reasoning behind subversion would be detectable.

For the first stage of Apollo’s assessment, we’ll provide access to selected public and unreleased models and training checkpoints, relevant training data and evaluation results, training intervention records, monitoring systems and flagged behaviors, and researchers who can provide context, alongside the tools and compute needed to conduct independent testing.

Goodfire

Goodfire will embed interpretability researchers and use the Goodfire platform to analyze model internals. The main objective of their assessment is training run monitoring, debugging, and alignment. The primary audience for their assessments are internal OpenAI researchers, who will use their findings to make safety improvements. Goodfire will also publish learnings and findings as the collaboration progresses.

METR

We intend to conduct two assessments with METR, focused on model alignment and monitoring:

  • Alignment assessment: For the first phase, METR will assess misaligned behavior in our largest reinforcement learning run to date, including failures that our existing tests and monitors may have missed. Within the selected run and agreed access, resources, and safety arrangements, METR would independently set assessment priorities and iterate on searches, analyses, and elicitation experiments. Access will include sampling relevant checkpoints, training and eval rollouts, and misalignment monitoring records for this run. For this first phase, the assessment will focus on our largest frontier RL run. [^1]
  • Monitor assessment: We recently worked with METR to conduct a monitor stress testing exercise. [^2] We have preliminarily agreed to conduct another monitor assessment which we intend to be broader, but this is pending mutual agreement on the assessment scope.

Transluce

Transluce will conduct an initial assessment of our post-training alignment evaluations, examining their quality and coverage. For the initial assessment, we will provide Transluce with access to selected public and unreleased models, our post-training alignment evaluations and relevant results, supporting model behavior data and safety case documentation, and researchers who can provide context, alongside the tools and compute needed to conduct independent testing.

OpenAI and Transluce also share the intent to conduct a follow-on assessment of approximately eight weeks in early 2027, with broader access to additional models, deeper access to our alignment evaluations, and information on how evaluation results inform model release decisions.

Protocols for embedded assessments

We’ve developed protocols and processes with input from the assessors to balance deep access with security, confidentiality, oversight, and the independence and integrity of assessments. We are sharing them to shed light on how we will work with embedded assessors, and to advance discussion about norms for embedded assessments, which have yet to be standardized. We expect that these will need to evolve as we learn more about how to best conduct these assessments. Especially in the early stages of embedded assessments, we expect technical access challenges, space constraints, evolving security protocols, and other unexpected issues to arise. We hope to work through these challenges collaboratively with our embedded assessors and share lessons learned.

Scoping and iteration

Each engagement begins with an agreed scope and the access, resources, and safety arrangements needed to conduct a rigorous assessment. Some assessments test specific safety claims; others give assessors room to discover unexpected failure modes within a defined system or training run. In both cases, assessors need the freedom to follow evidence, refine their methods, and propose further assessments when findings warrant it. We also have a process for addressing additional access needs as assessments develop. When direct access could expose our most sensitive IP, such as model weights or architecture, we will consider alternatives, including pairing assessors with OpenAI researchers to run analyses.

A secure, on-site environment

Assessors who need access to sensitive confidential information will work on site at OpenAI offices and be provided OpenAI devices to conduct their work. The primary purpose is to protect the security of the information they access. Working onsite also gives assessors the opportunity to engage with dedicated staff who can help them navigate their access and provide context for their assessments. We also expect our embedded assessors to follow at minimum [^3] the same security standards as OpenAI employees when running experiments and evaluations.

Oversight and dispute resolution

Our Safety and Security Committee (SSC), an independent oversight committee of the OpenAI Foundation’s Board, will have visibility into the embedded assessment and the ability to exercise their oversight responsibilities before, during, and at their conclusion. They will receive unredacted reports, and both the assessors and the SSC will be able to speak with each other during the assessment, including privately. Disagreements among OpenAI and any evaluator will be escalated to the Safety Advisory Group and OpenAI executives, who will have a specified time period to resolve them, and the SSC will be notified of any such disagreements promptly when escalated. If a disagreement remains, the matter will be referred to the SSC for resolution.

Employee discussions and attestation

Discussions with employees help assessors understand internal information and the context behind the work. OpenAI will authorize and provide guidance to staff members who can speak to assessors, and assessors may request additional staff to speak with as needed. To encourage open, candid conversations and protect the privacy of our staff, these discussions are off the record, meaning that they can inform the assessments but cannot be published or attributed to employees. Through a separate attestation pathway, assessors can request OpenAI-attested statements verifying, correcting, or clarifying something they learned in conversation with employees. These statements are not attributed to individuals or intended to characterize individual views. Assessors may publish statements that OpenAI attests to, or disclose that OpenAI declined to attest to a statement.

Conflict disclosures & cool off periods

Managing conflicts of interest clearly and transparently—including those involving personal relationships, financial interests, [^4] and other issues—is important to the integrity and independence of evaluations. Cool-off periods [^5] also need to be designed carefully, so assessors can work responsibly across labs while handling competitive information sensitively. We are working with our assessors to develop clear processes for both conflicts and cool-off periods, including allowing for flexibility in mitigations and transparency in decisions to request recusal.

Publication and responsible disclosure periods

Assessors maintain independence in publishing their assessments, but we will retain the right to redact confidential information, consistent with our previous third-party engagements. We may also request responsible disclosure delays to allow time to act on findings that could pose security or safety risks and remediate issues that remain.

Public reporting will help people understand the findings and their limitations, while more detailed confidential reports give oversight bodies the evidence they need to act. We want these assessments to improve how we identify and address risks, and give the public a clearer basis for judging our safety work.

Looking ahead

These assessments bring an independent perspective to our ongoing alignment and safety work. Public reporting on their findings will make that work more transparent and provide greater assurance about how risks are being identified and addressed. We expect this first phase of assessments to lead to longer-term, consistent embedded engagements over time that can continuously provide third party safety signals and recommended improvements on our most load bearing safety claims and safety cases.

We’ll continue working with multiple independent third parties to help build a healthy assessment ecosystem with diverse expertise and approaches. Some investigations need sustained, on-site work with sensitive systems; others primarily evaluate deployed or soon to be deployed models or review public documentation.

In the coming months, we’ll support external assessments across a broader range of safety questions. We are also exploring future embedded assessments that would examine compliance with our safety frameworks, as well as our organizational culture and processes. We are committed to expanding this work and helping build the capacity and shared standards needed to sustain it.

Footnotes

  1. This scope is less expansive than what METR outlines in their alignment assessment proposal, which envisions a broad survey of misalignment.

  2. We recently conducted a 3 week long monitor red teaming exercise with METR, which granted them access to our codebase, monitors, relevant trajectories, and discussions with OpenAI research staff for the purpose of better understanding our monitors. https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/

  3. In some cases, there will be more stringent standards applied depending on the conditions of the experiment to allow for careful security vetting.

  4. Apollo Research and Goodfire accept financial compensation for their work, but their compensation is not contingent on the results of their assessment. METR and Transluce do not accept financial compensation for their assessments, but may accept API credits / compute to use within their assessments. We do not believe that accepting financial compensation prevents assessors from maintaining independence - but we believe it is important to disclose this.

  5. We ask staff at third party organizations who participate in an embedded assessment to observe a brief cool-off period before embedding with another AI developer, to reduce the risk of inadvertent disclosure of confidential information.