Pyyan / News / 16 November 2026

SafetyOpenAI · Apollo Research · METR · Transluce · Goodfire

OpenAI signed Apollo Research, METR, Transluce and Goodfire for embedded safety assessments

4independent assessors in the first phase, with access to internal models and data

OpenAI announced its first embedded assessments on 16 November. Four outside groups get deep access to internal models and systems. Apollo Research will test whether models could covertly work against their developers. METR will look for misaligned behaviour in OpenAI's largest reinforcement learning run. Transluce will review OpenAI's post-training alignment evaluations. Goodfire will study model internals during training. Assessors handling sensitive data work on site, and they can publish with confidential details redacted. OpenAI says this supports its White House Accord commitments.

OpenAICNBCfrom the source itself

Related

← All the news, newest first