OpenAI has signed agreements with four independent organizations, Apollo Research, METR, Transluce and Goodfire, for the first phase of a program it calls embedded assessments, the company announced on November 16, 2026. The groups will review the company's models and safety work from inside OpenAI instead of only testing finished models before release. (The Neuron (opens in a new tab); Superpower Daily (opens in a new tab))

According to The Neuron, the researchers receive access to selected internal models, training records and safety systems, and some of them can interview OpenAI employees. 36Kr, citing OpenAI president Greg Brockman, described it as the company's first round of embedded assessments. (The Neuron (opens in a new tab); 36Kr (opens in a new tab))

What each group will study

Apollo Research is sending six researchers to work inside OpenAI for four months. Their questions are whether training creates incentives for models to undermine the safeguards meant to keep them under human control, whether safety tests can expose that behavior, whether it shows up when models are used internally, and whether the reasoning behind it can be detected. For the training work, Apollo is promised access to code, datasets, reinforcement learning records and saved model versions, Superpower Daily reported. (The Neuron (opens in a new tab); Superpower Daily (opens in a new tab))

METR will examine OpenAI's largest reinforcement learning training run to date, looking for behavior that conflicts with the goals developers intended. Goodfire, which works on interpretability, will study what happens inside models during training to explain the processes behind their behavior. Transluce will spend six weeks testing whether OpenAI's safety evaluations can catch deception, manipulation or attempts to game scoring. (The Neuron (opens in a new tab); 36Kr (opens in a new tab))

Limits on access and publication

OpenAI agreed that it cannot withhold findings simply because they reflect badly on the company. It can restrict sensitive material for technical, security or intellectual property reasons, but it must say what it withholds and why. It can also redact confidential evidence from public reports, while full reports go to the Safety and Security Committee of its board. (The Neuron (opens in a new tab); Superpower Daily (opens in a new tab))

Apollo plans two public reports, and OpenAI may redact evidence but not Apollo's conclusions. Apollo also secured what it calls meta-transparency, the right to describe access limits, redactions and any pressure it believes shaped its work without prior review by OpenAI. Transluce can discuss findings about released models but not about models used only internally. (The Neuron (opens in a new tab); Superpower Daily (opens in a new tab))

The terms leave OpenAI with real influence. The company chooses which staff Apollo may speak with, although Apollo can suggest names, and claims drawn from employee conversations need OpenAI's confirmation before publication unless Apollo verifies them independently. Apollo acknowledged that the agreement does not yet meet every requirement for a fully independent evaluation. Apollo and Goodfire are paid, with payment not tied to their findings, while METR and Transluce take no money but may receive computing resources or API credits. (The Neuron (opens in a new tab); Superpower Daily (opens in a new tab))

The program follows earlier outside work at the company. In October, METR and Redwood Research found weaknesses in an earlier version of OpenAI's agent monitoring system, including one that could keep hundreds of recent agent actions from reaching the monitor. OpenAI said most were patched, but METR has not independently verified the fixes, The Neuron reported. (The Neuron (opens in a new tab))

In NeuraFeed's reading, the agreement moves outside review closer to the point where problems are created, inside training, rather than at launch. Its value will depend on how much the public reports can show after redactions, and on whether the auditors use their right to describe what they were not allowed to see.