OpenAI internal agent alignment monitoring evaluation report made public; participant says publication was edited and delayed by two and a half months
Two employees from METR and Redwood Research spent three weeks embedded at OpenAI this summer, conducting an embedded evaluation pilot to assess OpenAI's system for monitoring internal agent alignment issues; their work brief is now public. The author who participated in the project said this was one of the third-party collaborations he took part in at OpenAI; he declined to comment on the scale of the edits OpenAI did not allow METR to make to the report, and expressed disappointment that there was no explanation for the two-and-a-half-month delay in publication.
This was one of the third-party engagements I worked on at OpenAI. I'm glad the report is out! I'm disappointed that METR wasn't allowed to comment on the scale of redactions made by OpenAI and that there is no explanation why the publication was delayed for 2.5 months.
Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.View the quoted post on X