What was announced?

OpenAI said it has started the first phase of its embedded assessments program. As part of the effort, the company will work with outside, independent organizations.

Who is involved?

Four organizations are named in the announcement: Apollo Research, METR, Transluce and Goodfire. These teams will examine OpenAI models independently of the company's own staff.

What will be assessed?

The evaluations cover three areas:

  • Model alignment: how closely a model behaves as intended.
  • Safeguards: how resistant a model is to harmful or unwanted outputs.
  • Monitoring systems: how well model behavior can be tracked.

Why it matters

A common criticism in AI safety debates is that evaluations stay inside the company that built the model. OpenAI's move brings independent experts directly into the process. The company describes this as a "first phase," suggesting the program may expand later.