AI safety researchers inside a frontier lab stress-test models before and after deployment — probing for dangerous outputs and measuring whether behavior matches stated values. Some focus on alignment; others design evaluations to detect whether a model could help plan a cyberattack or deceive its operators.

The friction is structural. A safety team's job is to find reasons to slow down or change a model; a product team's job is to ship. When findings conflict with release timelines, whose concerns get acted on — versus quietly filed — is exactly what critics argue about. Former OpenAI safety co-lead Jan Leike wrote on departure in 2024 that "safety culture and processes have taken a backseat to shiny products."

This week, OpenAI fired three safety researchers for allegedly sharing confidential information with an outside AI safety organization — reigniting the debate about whether safety teams inside commercial labs have any real power to slow things down.