OpenAI said on Nov. 16 it signed agreements with four independent assessors1 — Apollo Research, METR, Transluce and Goodfire — for a first phase of embedded safety assessments.

The move matters because OpenAI said the assessors will get deep access to internal models, checkpoints, monitoring records and other non-public materials, a setup meant to test alignment and safeguard claims that ordinary external benchmark access would not cover OpenAI1.

According to OpenAI1, the program focuses on "alignment and safeguards for models during training and internal use" and is intended to complement the company's safety cases. OpenAI said the assessors may receive access to model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation and discussions with researchers, depending on each evaluation's needs OpenAI1.

OpenAI's announcement lays out different scopes for each assessor. Apollo Research1 will examine whether models could "covertly work against their developers' intentions" and whether OpenAI's safeguards would detect that behavior. OpenAI said Apollo's first stage will cover selected public and unreleased models, training checkpoints, relevant training data and evaluation results, training intervention records, monitoring systems and flagged behaviors OpenAI1.

Goodfire1 will embed interpretability researchers to analyze model internals, with OpenAI saying the main objective is training-run monitoring, debugging and alignment. OpenAI said Goodfire will also publish learnings and findings as the collaboration progresses OpenAI1.

For METR1, OpenAI said it intends to conduct two assessments focused on alignment and monitoring. The first phase will assess "misaligned behavior in our largest reinforcement learning run to date," including failures OpenAI's existing tests and monitors may have missed, with access to checkpoints, training and evaluation rollouts, and monitoring records for that run. OpenAI also said a broader follow-on monitor assessment is only preliminarily agreed and remains pending mutual agreement on scope OpenAI1.

Transluce1 will review OpenAI's post-training alignment evaluations, including their quality and coverage, with access to selected public and unreleased models, evaluation results, behavior data and safety case documentation, according to OpenAI. OpenAI said the two sides also share an intent to run a follow-on assessment of about eight weeks in early 2027 with broader access OpenAI1.

A separate post by Transluce2 gives additional terms for its own pilot, but not independent confirmation of the overall OpenAI announcement. Transluce said its initial exercise is a six-week pilot starting in early December, with onsite work at OpenAI and access to "a handful of OpenAI models," including at least one internal-only model and raw chain-of-thought outputs Transluce2.

Transluce also said it is not accepting payment from OpenAI for the pilot, while OpenAI will cover compute, devices, infrastructure and temporary workspace Transluce2. Transluce said OpenAI can request redactions of intellectual property and other non-public information before publication, and that the agreement bars Transluce from publishing results related to internal-only models Transluce2.

Access, publication and oversight terms

OpenAI said assessors who need sensitive confidential information will work on site at OpenAI offices using OpenAI devices, and that the company expects embedded assessors to follow at minimum the same security standards as employees OpenAI1. The company said its Safety and Security Committee, which it describes as an independent oversight committee of the OpenAI Foundation's Board, will receive unredacted reports and can speak privately with assessors during the process OpenAI1.

OpenAI also said assessors maintain independence in publishing their assessments, but that the company will retain the right to redact confidential information and may request disclosure delays to address findings that could pose security or safety risks OpenAI1.

Those terms define what these agreements do and do not show. The Nov. 16 post is an issuer announcement describing planned assessment scopes and process. It does not include any completed assessment results.

Broader safety push

The company framed the embedded assessments as part of a wider safety effort, saying they build on earlier work with third-party evaluators and support commitments under the White House Accord on Super Intelligence OpenAI1.

Automatica reported on Nov. 13 that OpenAI funded a $2 million children's AI safety research network3, another recent sign of the company's effort to tie public safety commitments to outside research partnerships. That earlier report is background, not independent confirmation of the new agreements.

OpenAI said it expects this first phase to lead to longer-term embedded engagements that can provide continuing third-party safety signals and recommended improvements on its safety claims and safety cases OpenAI1.