OpenAI Launches Embedded Safety Assessments Program

Four outside research groups will gain access to OpenAI models, checkpoints and internal safety tooling under the first phase of the program.

Updated on Nov. 16, 2026 in Artificial Intelligence

Isometric editorial illustration of a plain slate-gray access panel with a keyhole recess set into a pale concrete wall, evoking controlled outside access to guarded systems.
OpenAI launched an embedded safety assessments program giving four outside research groups, including Apollo Research, Goodfire, METR and Transluce, access to its models, checkpoints and internal safety tooling. AI Illustration. Upload story photo >

Live Poll

Do you trust AI companies to test their own systems for safety without outside oversight?

OpenAI launched the first phase of its embedded assessments program with four outside research groups: Apollo Research, Goodfire, METR and Transluce. The program gives these researchers access to model checkpoints, source code, training and testing datasets, and monitoring infrastructure.

Why it matters

The initiative gives independent researchers a closer look at OpenAI's internal safety work, which is intended to help surface risks such as models acting against their developers' intentions. Conflict-of-interest disclosures, cooling-off periods and compensation structures are designed to limit the influence of the outside groups on OpenAI.

The program involves four outside research groups, and the planned follow-up project runs roughly eight weeks in early 2027. Findings have not yet been released, and OpenAI has said they will be shared publicly eventually.

The players

OpenAI

OpenAI is an artificial intelligence research company that develops large language models and related products, and it designed the embedded assessments program to give outside researchers closer access to its internal safety work.

Apollo Research

Apollo Research is an AI safety organization that studies whether advanced models may pursue hidden goals or deceive their developers, and it will examine whether OpenAI models could secretly act against their developers' intentions.

Goodfire

Goodfire is an AI interpretability company that works on understanding how neural networks represent information internally, and it will focus on examining how OpenAI's models work internally.

METR

METR is a research nonprofit that evaluates frontier AI models for dangerous capabilities and misaligned behavior, and it will look for potentially misaligned behavior during OpenAI's largest reinforcement-learning training run to date.

Transluce

Transluce is an AI research organization focused on understanding and evaluating the behavior of language models, and it will review OpenAI's post-training alignment evaluations.

The details

Apollo Research will examine whether OpenAI models could secretly act against their developers' intentions, while Goodfire will focus on interpretability. METR will look for potentially misaligned behavior during OpenAI's largest reinforcement-learning training run to date, and Transluce will review OpenAI's post-training alignment evaluations. Researchers may use OpenAI-provided devices for on-site access to sensitive materials, and discussions with OpenAI staff will remain off the record.

Timeline

  1. November 16, 2026: OpenAI announced the first phase of the embedded assessments program.

  2. Early 2027: A follow-up eight-week project with broader access is planned to begin.

Roadmap

The program extends the practice of bringing outside evaluators such as METR and Apollo Research into frontier-model safety testing, which has largely relied on pre-deployment evaluations. Granting access to checkpoints, source code and training data marks a deeper level of external scrutiny than those earlier arrangements typically provided.

For users of OpenAI's products, the program is aimed at identifying risks such as models acting against developers' intentions before they reach the public. Because findings will be shared publicly eventually, users may later see more detail on how these safety checks were conducted.

The takeaway

Outside evaluation of AI systems is becoming more formalized, with deeper access to internal materials than before. Readers following AI safety should watch for the eventual public release of the program's findings.

What happens next

A follow-up project lasting roughly eight weeks, with broader access for the outside researchers, is planned to begin in early 2027.

Further reading

For more on how outside groups are testing frontier systems, see the Artificial Intelligence section.

Live Poll

Do you trust AI companies to test their own systems for safety without outside oversight?