OpenAI Gives Internal Models to Four Organizations for Review, but the Comment Section Only Asks Where 6.2 Is
OpenAI announces its first batch of embedded assessments: Apollo Research, METR, Transluce, and Goodfire gain access to internal models, code, training data, and monitoring systems. The assessment protocol is spelled out in great detail, while the comment section is pressing for a new model.
On November 16, OpenAI announced its first batch of embedded assessment partnerships. The four independent organizations received not only public interfaces, but also internal model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and opportunities to speak directly with frontline researchers and engineers.
The partners are Apollo Research, METR, Transluce, and Goodfire. OpenAI says this type of assessment allows outside experts to work deeply within internal systems for a period of time to examine model alignment and whether safety protections are effective. The details were published on the OpenAI website.
What Each of the Four Organizations Examines
Apollo Research will assess whether models secretly act against developers’ intent and whether existing safeguards can detect such behavior. It will examine evidence from training, safety testing, and internal use, and verify four claims in two phases: that training does not incentivize subversion; that strong evaluations and red-teaming do not elicit subversion; that models do not attempt subversion during internal use; and that the reasoning behind subversion can be detected.
METR’s task focuses on model alignment and monitoring. In the first phase, METR will assess misaligned behavior in OpenAI’s largest reinforcement-learning training run to date, including failures that existing tests and monitoring might miss. OpenAI will also provide relevant checkpoints, training and evaluation rollout records, and misalignment monitoring records. The two sides have also tentatively agreed to conduct a broader monitoring assessment, though its scope is still under discussion.
Transluce will assess alignment evaluations in the post-training phase, focusing on quality and coverage. It will develop additional tests and propose improvements for testing blind spots. The two sides plan to conduct a follow-up assessment of about eight weeks in early 2027, with access to more models, and to understand how evaluation results influence model release decisions.
Goodfire’s interpretability researchers will use their own platform to analyze model internals, with goals including training-run monitoring, debugging, and alignment, helping OpenAI researchers identify and improve safety issues.
The Protocol Is Spelled Out in Detail
Access is broad, and so are the restrictions. Evaluators who need access to sensitive confidential information must work on-site at OpenAI offices and use OpenAI-provided equipment. OpenAI says this is mainly to protect information security and to make it easier for evaluators to communicate with dedicated staff. When running experiments, evaluators must follow at least the same safety standards as OpenAI employees.
On oversight, the Safety and Security Committee (SSC) of the OpenAI Foundation Board can intervene before, during, and after the assessment, receive unredacted reports, and speak privately with evaluators. If OpenAI and evaluators disagree, the issue will first be escalated to the safety advisory group and executives for resolution within a set time; if it remains unresolved, it goes to the SSC.
Discussions with employees are off the record. Evaluators can use them to form judgments, but may not publish them publicly or attribute them to specific employees. There is also an attestation pathway: evaluators can ask OpenAI to attest to a statement in order to verify, correct, or clarify information learned from employees. OpenAI may attest or decline, and evaluators may publish the attestation statement or disclose that OpenAI declined to attest.
Conflict of interest and cooling-off periods are also written into the protocol. Apollo Research and Goodfire receive financial compensation, but the compensation is not tied to assessment results. METR and Transluce do not accept financial compensation and may accept API credits or compute. Third-party employees participating in embedded assessments must observe a short cooling-off period before embedding with another AI developer, reducing the risk of confidential information leaks.
On publication, evaluators remain independent, but OpenAI reserves the right to redact confidential information and may request a delay for responsible disclosure in order to address issues that could pose safety risks.
The Comment Section Is Pressing for the Model
Reactions under the post are polarized. Some ask whether the four organizations have already received GPT-6.2 Astra. Some ask whether it can be finished before Thursday. Some simply say they want the new model.
Some users call safety assessments a new excuse for delaying releases, saying Anthropic also takes safety seriously yet can still release frontier models. Others say they understand, believing these steps should be taken even if they slow the pace.
A reposted reply video captures this sentiment well:
Based on public information, OpenAI has this time moved external assessment from 'looking at public materials' to 'going into internal systems.' The assessment scope, access boundaries, oversight pathways, and disclosure rules are all spelled out. What the comment section cares about, however, is something else: when the model will be released. The tension between safety work and product cadence is laid out clearly under this tweet.
Original link: https://openai.com/index/our-first-phase-of-embedded-assessments/
发布时间: 2026-11-17 03:28