Our updated Preparedness Framework

Source: OpenAI Date: 2025-04-15 URL: https://openai.com/index/updating-our-preparedness-framework

Summary

Summary

OpenAI published an updated version of its Preparedness Framework in April 2025 — revising the risk thresholds, evaluation methodologies, and governance processes for assessing frontier AI systems before deployment. The update reflected lessons from applying the original framework (published in late 2023) to o1, o3, and GPT-5 evaluations.

Implications

Safety/governance thread. The Preparedness Framework update is significant because it documents how OpenAI’s safety evaluation process evolved in response to experience with more capable models. The original framework established the concept of “critical” risk thresholds that would pause deployment — the update presumably revised what those thresholds mean as models got more capable and evaluation methodology matured. The April 2025 timing is between the o3/o4-mini launch and GPT-5’s launch a month later, suggesting the update was partly motivated by evaluations that surfaced in the o3 and GPT-5 pre-deployment work. The framework’s credibility depends on whether the stated thresholds are actually binding, which remains difficult to verify externally.