Source: LessWrong
OpenAI's latest model underperforms on safety evaluation tasks that would typically be public-facing commitments—a gap between marketed capabilities and measured safety performance. This raises questions about both the rigor of their evaluation standards and transparency to external oversight bodies. Safety claims have become central to competitive and regulatory positioning for frontier labs. If performance lags rhetoric, it signals either evaluation methodology problems or that safety optimization is deprioritized relative to capability scaling. The finding surfaced through independent research rather than disclosure, suggesting structural misalignment between OpenAI's public safety commitments and internal development incentives.