OpenAI signed agreements with four independent groups to test its models from inside the company. Six researchers from Apollo Research, a safety lab, start a four-month stay this week. They get unreleased models and training data, while OpenAI can still remove confidential details from their reports.