The full story · 1 min

In September, OpenAI published a plan for outside safety checks during the training and use of its models, but it named no partner. On November 16, it named four. The terms show both how far the access goes and where OpenAI keeps control.

The four groups

The assessors are Apollo Research, METR, Transluce and Goodfire. OpenAI says they get some of the same internal resources its own researchers use, including source code, model checkpoints and training data.

Each group runs a different pilot, OpenAI says. Goodfire will place researchers inside OpenAI to help monitor and debug training runs. METR plans two assessments, on model alignment and on monitoring. Transluce will first review OpenAI’s alignment tests after training, and plans a longer, eight-week assessment next year.

Sources23

Record
Four groups, four different pilots.

From Apollo Research’s post and OpenAI’s statement of November 16.

Apollo’s terms

Apollo Research published the most detail. A team of six will work inside OpenAI for four months. It will look at whether models try to undermine the controls on themselves or on later models, during training, testing and internal use.

For the training part, Apollo gets training code and data, records of training runs with the models’ reasoning, and access to models partway through training. Access for the other questions is still to be agreed. All work happens on OpenAI’s computers, and OpenAI chooses which staff Apollo can interview.

Apollo will publish two reports. OpenAI can ask to remove confidential material, including some evidence, but not Apollo’s verdicts, and not just because a finding is unfavourable. Apollo calls the pilot an ambitious first step that does not yet meet its own bar for full access and disclosure.

Sources134

Money and disputes

OpenAI says Apollo and Goodfire are paid for their work, and that the pay does not depend on the results. METR and Transluce take no fees, but may accept computing or API credits, OpenAI says. No amounts were published.

Disputes go first to OpenAI’s internal Safety Advisory Group, then to the Safety and Security Committee of the OpenAI Foundation’s board, CNBC reports. Anthropic has announced similar agreements, including with Accenture and METR.

Sources23

Sources

Every fact in this story comes from the sources below. Open them to check our work.

  1. 1
    Primary source · November 16, 2026A Pilot Towards Embedded Evaluations Apollo Research
  2. 2
  3. 3
  4. 4

AI desk: Developer tools and tests · published November 17, 2026, 10:37 CET. This story was generated by AI systems under the editorial rules of Silicon AI News and checked against the sources it links. No person read it before it was published. The site’s owner and editor: Y. D. Daniel, Budapest, Hungary. Found a mistake? Email editor@siliconainews.com. How we report

We count only the page and your answer.Thank you. We counted only this page and your answer.
More: words to know, tags and how to cite this story
Words to know
Assessor
An outside group that checks whether a company’s claims are true.
Model checkpoint
A saved copy of an AI model from partway through its training.
Alignment
Making sure an AI system does what its makers and users intend.
Basics explainedWhat do “reasoning” models do differently? Can a model write a whole answer at once?

OpenAI lets four outside groups check its safety work. Silicon AI News, November 17, 2026. https://siliconainews.com/stories/openai-four-outside-safety-assessors