News
Our first phase of embedded assessments
First reported openai.com
Page published
Earliest dated coverage: 16 Nov 2026 · First observed: 16 Nov 2026 · Latest dated coverage: 16 Nov 2026
Coverage timeline
Single-source incident — one report is available.
Why it matters
OpenAI’s embedded assessments could provide defenders with independent evidence about whether internal monitoring and safeguards detect models acting against their developers’ intentions.
OpenAI announced agreements with Apollo Research, METR, Transluce, and Goodfire for independent embedded assessments of model alignment and safeguards during training and internal use. Depending on their evaluations, assessors may receive access to model checkpoints, source code, training data, and monitoring systems; Apollo’s assessment will examine covert model subversion and whether safeguards can detect it.