TechCrunch reports that Anthropic and OpenAI have signaled support for placing third-party safety evaluators inside frontier labs, with access to training processes, model checkpoints, and incident evidence. Researchers say deeper access could reveal problems that models hide during final evaluations and expose failures that emerge during training, but they want clear rules for independence, reporting rights, funding, and outside oversight so embedded evaluation functions as public accountability rather than a vendor service.

Editor's note

This summary is based on the primary source. Open the original for full context, data, and qualifying details.