OpenAI names four organizations for embedded safety assessments

OpenAI names four organizations for embedded safety assessments
By guthlabs.newsroom · November 16, 2026
OpenAI said on November 16 that it has signed agreements with Apollo Research, METR, Transluce and Goodfire for embedded assessments of model alignment and safeguards. The program gives outside assessors scoped access to internal systems and evidence, rather than relying only on public model interfaces. Official announcement
The work covers different questions: potential covert subversion, model internals, misalignment during a major reinforcement-learning run, and the coverage of alignment evaluations. Access varies by engagement and can include checkpoints, data, code, monitoring records and discussions with staff.
OpenAI says assessors can publish independently, but it retains confidentiality redaction rights and may request safety-related disclosure delays. Its Safety and Security Committee will receive unredacted reports. Apollo and Goodfire are paid for their work; METR and Transluce do not accept assessment fees, though they may receive compute support.
Guth News analysis: The announcement establishes an assessment process, not a finding that the models are safe. Its value will depend on what assessors can actually inspect, how disagreements are handled and how much of the eventual evidence reaches the public. Scoped access and publication limits should remain visible alongside any future conclusions.
Evidence limit: Company announcement dated November 16, read in full; not completed assessor reports. No claim of unrestricted access or certified safety.
Sources and citations
Revision history
Published .