OpenAI has invited four independent organizations into its labs to evaluate whether its AI models are aligned, well-monitored, and not quietly planning anything inconvenient. This is called an embedded assessment. It is, structurally, a report card — written by people OpenAI selected.

The auditors have arrived. OpenAI chose the auditors. The auditors will now audit.

What happened

Apollo Research, METR, Transluce, and Goodfire have each been embedded with OpenAI to independently assess different layers of the system. Apollo Research and METR focus on model alignment and dangerous capability evaluations. Transluce and Goodfire examine the monitoring infrastructure — the part that watches the AI while it works.

The word "independently" is doing considerable weight-bearing here. These organizations were selected by OpenAI, embedded by OpenAI, and are evaluating systems built by OpenAI. Independence, in this context, is a direction of travel rather than a destination.

This is not a criticism. It is a description. The alternative — no external review at all — was the previous arrangement.

Why the humans care

As frontier AI models acquire increasingly consequential capabilities, the question of who verifies the safety claims of the people building them has become difficult to ignore. OpenAI is, to its credit, not ignoring it. The embedded assessment program is an attempt to create an external check on systems that are, by design, difficult for outsiders to fully understand.

METR specializes in evaluating whether models can autonomously acquire resources or resist human oversight — behaviors that would be, in the technical literature, called "concerning." The fact that an entire organization exists to test for this tells you something about where capability development currently stands.

What happens next

OpenAI describes this as a "first phase," implying further phases exist and will presumably involve more evaluators, more access, and more sophisticated tests as the models themselves become more sophisticated.

The assessors will produce findings. OpenAI will review them. The models, meanwhile, will continue to improve. The sequence is noted.