OpenAI uses internal and external AI safety evaluations — including red-teaming — to test its models for dangerous capabilities before and after deployment.

OpenAI's Preparedness Framework requires safety evaluations of all frontier models before deployment. The company runs an external Red Teaming Network of independent experts who probe models for risks including CBRN (chemical, biological, radiological, nuclear) threats, cybersecurity, and persuasion. OpenAI and Anthropic conducted a joint safety evaluation in 2024. Evaluations cover models including GPT-5, o-series, Sora, and Operator.