OpenAI has moved safety oversight from a centralized team to decentralized product departments. For your business, this means you must increase your own independent testing and human oversight when deploying AI, as vendor-provided safety assurances may become less centralized and more integrated into commercial release cycles.
OpenAI has dissolved its dedicated preparedness team, shifting safety oversight to product teams. Learn how this affects AI security and your business risk.
OpenAI disbanded its dedicated preparedness team, which was previously tasked with identifying catastrophic risks in AI models. The responsibilities previously held by this team have been redistributed to existing product and research units to integrate safety assessments directly into the development cycle.
Whether this change impacts safety is debated. While OpenAI argues that distributed safety leads to more practical and timely mitigations, critics suggest that removing an independent, centralized team reduces the necessary friction that prevents the premature release of potentially risky or unvetted technology.
Businesses should implement independent red-teaming and human-in-the-loop validation for all AI deployments. Because model providers are shifting their internal safety structures, you should not rely exclusively on vendor safety claims and instead perform your own risk assessments tailored to your specific use cases.
Red-teaming involves intentionally testing a system for vulnerabilities, biases, or dangerous capabilities. It is a critical component of AI safety that identifies how a model might be exploited or malfunction before it is made available to the public or integrated into sensitive workflows.