OpenAI announced it will integrate third-party organizations into the early stages of AI model training, evaluation, and deployment for technical safety assessments. Previously, external teams were typically engaged only before a model's release to evaluate capabilities and risks. The company stated that these evaluations must remain independent, scientifically rigorous, and include strong security mechanisms with clear divisions of responsibility. OpenAI is currently in discussions with multiple new and existing partners, including METR and Redwood Research, to expand this initiative. Some evaluators may be placed inside the company's offices to participate in testing during the most sensitive development stages, marking a shift toward continuous external oversight throughout the model lifecycle.