OpenAI’s Safety and Security Committee, constituted in May 2024 as a board-level oversight mechanism following the November 2023 CEO crisis, has been reorganized again. The structure now seats safety evaluation primarily within the technical organization rather than as an independent board-level function, a shift from the original intent.
The pattern across the past 18 months at OpenAI, and to varying degrees at other frontier labs, is legible: governance structures created to provide independent oversight tend to migrate toward advisory roles as commercial velocity increases. The tension is structural — truly independent safety oversight that can block or delay a product release has a real commercial cost, and organizations operating in a competitive market feel that cost acutely.
This is not a claim about any specific individual or decision at OpenAI. It’s an observation about organizational dynamics. The structure of AI safety governance at frontier labs reflects who actually bears the downside risk of moving too fast versus too slow. Current structures put shareholders and management in the downside-risk-of-too-slow category and distribute the downside-risk-of-too-fast across a broader and less powerful constituency.
The policy relevance: third-party evaluation, external red-teaming requirements, and government-mandated disclosure frameworks gain importance precisely because internal governance has shown its limits. The EU AI Act’s third-party audit requirements for high-risk systems reflect this analysis. The US government’s approach — voluntary commitments with executive order follow-on — reflects a different political economy.
For enterprise AI buyers: the governance signals at frontier labs are relevant to vendor risk assessment. Organizations building on AI infrastructure should understand the governance structure of their key providers and factor that into business continuity planning.