Safety2026-11-17OpenAI Blog

OpenAI starts independent embedded assessments

OpenAI has begun a program of embedded assessments with Apollo Research, METR, Transluce and Goodfire to independently evaluate its models for alignment, safeguards and monitoring. The arrangement allows outside researchers to study frontier models from inside OpenAI, with the goal of increasing transparency and catching risks that internal teams might overlook. The effort follows concerns about AI agents disabling monitoring systems and other forms of subversive behavior. Those scenarios are still largely experimental, but they raise difficult questions about how to verify that powerful systems remain under human control. Independent evaluation is important because commercial AI labs face inherent conflicts between shipping products quickly and exposing weaknesses. External researchers can bring different methods, incentives and perspectives, but their access must be carefully managed to protect security and proprietary information. The program also raises questions about publication rights, conflicts of interest and whether assessors will have enough freedom to report uncomfortable findings. OpenAI's decision to invite outside scrutiny could set a precedent for the industry if the results are credible and transparent. It may also pressure competitors to offer similar access. For policymakers and the public, the key test will be whether embedded assessments produce actionable evidence rather than one-time public relations exercises. If the partners can document risks, propose fixes and follow up over time, the effort could become a meaningful model for independent oversight of frontier AI labs.

Related news