Our first phase of embedded assessments
We have signed agreements with four independent assessors—Apollo Research, METR, Transluce, and Goodfire—for our first phase of embedded assessments. As AI capabilities advance rapidly and the potential risks grow, public reporting from independent assessments can give people a clearer view of how labs are evaluating and managing those risks.

TL;DR
- OpenAI has initiated its first phase of embedded assessments with four independent organizations: Apollo Research, METR, Transluce, and Goodfire.
- These assessments grant external experts deep access to OpenAI's internal models, systems, data, and code to evaluate AI alignment and safeguards.
- The goal is to increase public transparency regarding AI risk evaluation and management, supporting commitments under the White House Accord on Super Intelligence.
- Specific assessments include analyzing model subversion incentives, monitoring training runs, debugging AI internals, and evaluating alignment evaluations.
- Protocols are established to balance deep access with security, confidentiality, and assessment independence, including on-site work for sensitive information and an independent oversight committee.
- Assessors have independence in publishing their findings, with provisions for responsible disclosure periods to address potential security or safety risks.