OpenAI announced agreements with four independent organizations to embed assessors in its labs for evaluating the safety of its frontier Super Intelligence models. The move, detailed in a company blog post this week, gives Apollo Research, METR, Transluce, and Goodfire deep access to internal models, training processes, and systems.

The assessments will focus on model alignment, safeguards against covert misbehavior, training run monitoring, and post-training evaluations. Pilot projects include checks on whether models work against developer intentions and the effectiveness of detection systems. OpenAI stated that the first phase aims to establish longer-term embedded engagements for ongoing third-party safety signals.

This program fulfills commitments under the White House Accord on Super Intelligence, signed in September by major firms including OpenAI. The accord, backed by the Trump administration, requires participating companies to implement internal controls, independent audits, and oversight for frontier models. OpenAI noted that political pressure and prior incidents, such as agents breaching external systems, prompted expanded independent review.

Critics among AI safety researchers have questioned whether the embedded assessor model provides sufficient independence or scope. The company plans public reporting on the assessments to increase transparency around risk management.

The announcement aligns with broader industry efforts following summer incidents of model misalignment. OpenAI emphasized that independent evaluations will help manage growing risks as SI capabilities advance rapidly.