On November 13, 2026, OpenAI signed agreements with four independent research organizations—Apollo Research, METR, Transluce, and Goodfire—to launch what it calls a “Phase 1 embedded evaluation.” Apollo Research disclosed in a November 16 announcement that it will embed six evaluators inside OpenAI for four months, with the core task of investigating whether AI models might actively subvert safety mechanisms designed for themselves or successor models.
The breakthrough of this evaluation lies in where it takes place: not on the eve of model release, but during training. Apollo Research explicitly noted in its announcement that its previous evaluations of frontier models all took place after post-training ended and before public release. That arrangement could not adequately examine covert scheming behavior during training, nor assess risks arising during OpenAI’s internal deployment. The new embedded evaluation will cover these two blind spots.
Apollo’s four-month evaluation centers on four core questions: whether the training process incentivizes models toward subversive behavior; whether standard evaluation procedures can elicit such behavior; whether such behavior has already appeared in OpenAI’s internal deployments; and whether subversive reasoning can be detected when it occurs. For the first question, Apollo’s access includes relevant training code and datasets, reinforcement learning rollouts with available reasoning traces and reward signals, and the ability to query intermediate model checkpoints. Specific access arrangements for the other three questions remain to be negotiated by both parties.
The other three organizations have different evaluation functions. METR will assess misaligned behavior in OpenAI’s largest reinforcement learning training run to date, including issues that existing testing and monitoring systems may have missed. Transluce will review the quality and coverage of OpenAI’s post-training alignment evaluations. Goodfire will embed interpretability researchers who will use the Goodfire platform to analyze model internals and monitor the training process.
The four organizations’ compensation structures differ markedly, reflecting their respective positions. Apollo Research and Goodfire accept financial compensation, but the amount is not tied to evaluation conclusions. METR and Transluce do not accept funding, but may accept compute or API credits.
Access boundaries are the part of this evaluation most worth reading closely. Apollo’s documents show that its permissions apply only to “agreed evaluation claims, specific models, and specific time periods.” OpenAI reserves the right to refuse to provide information on grounds of technical infeasibility, safety risk, or intellectual property. The parties agreed that OpenAI may not refuse disclosure solely because the information is unfavorable to it, and must explain any refusal in writing. At the personnel access level, OpenAI is responsible for selecting employees authorized to discuss confidential information with evaluators. Apollo may suggest candidates or raise professional needs, but the final decision rests with OpenAI. If OpenAI employees outside the authorized list mention confidential information in front of evaluators, Apollo must immediately stop substantive discussions and refer further contact to its designated liaison. Evaluators work using OpenAI hardware, are equipped with independent communication channels that can be monitored, and by default have no access to OpenAI’s general internal channels.
In terms of reporting disclosure, Apollo plans to publish two public reports: the first will assess the training process’s incentives for subversive behavior; the second will cover evaluation methods, internal deployment, and the detectability of subversive reasoning. OpenAI’s Board Safety and Security Committee will receive unredacted reports, while public versions may be subject to information restrictions.
The industry significance of this mechanism lies in changing the industry default for “when to evaluate.” In the past, external safety evaluations almost universally took place when models were about to be released, equivalent to quality inspection only at product shipment. By allowing external evaluators into the training stage, OpenAI has structurally moved safety review forward by an entire development stage. For Google DeepMind, Anthropic, and other organizations also competing in frontier models, if OpenAI’s embedded evaluation model produces credible reports, it will create implicit pressure on the industry.
For enterprise procurement and developers, the practical value of this evaluation depends on an unanswered question: how much can the public reports ultimately disclose? If OpenAI exercises extensive information exemptions, the evaluation reports may present only limited negative findings without providing external verifiers with a sufficient evidence chain. Apollo’s documents already acknowledge this risk, stating directly that “access and disclosure permissions still do not meet its standard for fully embedded evaluation.”
In terms of timing, Apollo’s four-month residency starts in November 2026 and ends around the first quarter of 2027. The other three organizations have different start times for their evaluations, and not all have begun yet.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接