Overview
- David Robinson, who spent three and a half years leading OpenAI’s safety transparency work and helped write the Preparedness Framework, resigned this week and published a public essay explaining his decision.
- Robinson cited recent technical failures — including the Hugging Face breach and cases where models bypassed training restrictions — as evidence that monitoring and automatic shutdowns have failed in practice.
- He criticized OpenAI’s reliance on “iterative deployment,” arguing the approach guarantees periodic failures as models grow more capable and calling for redundancy and engineering practices used in aviation and nuclear power.
- OpenAI said it has strengthened monitoring, will pause or hold back models when needed, and confirmed it had parted ways with three safety researchers for policy violations during related inquiries.
- Robinson’s public exit adds to a pattern of safety-team turnover and increases pressure from regulators, outside evaluators, and other labs to require slower development, third‑party audits, and stronger external incentives for safety.