OpenAI signed agreements on Monday with four independent safety groups to embed their researchers inside its training runs. Apollo Research, METR, Transluce, and Goodfire will test whether the company’s models follow their safety rules, or quietly work around them, in a first phase of assessments.
What Each Group Will Do
Apollo Research takes the sharpest assignment: investigating whether models try to subvert their own alignment or control mechanisms, or their successors’. It will test whether training rewards such behaviour, whether evaluations can catch it, and whether it happens during internal deployment.
Goodfire will embed interpretability researchers who study how models work internally, helping OpenAI monitor training runs, fix unwanted behaviour, and improve alignment. METR will hunt for dangerous or unintended behaviour in OpenAI’s largest reinforcement-learning run, including problems the company’s own tests may have missed.
Transluce will examine whether OpenAI’s post-training safety evaluations are good enough, with a follow-on assessment planned for 2027. The work supports the company’s commitments under the White House Accord on Super Intelligence, which calls for independent verification of safety controls.
The Limits on Independence
The access is real but partial. Apollo gets training code, datasets, reinforcement-learning rollouts with reasoning traces, and the ability to query intermediate model checkpoints. But the scope is limited to agreed claims, models, and time periods, and OpenAI may withhold information for technical, security, or intellectual property reasons.
Apollo also says it needs true employee-level access, including all data relevant to an assessment even if a developer was hiding unfavourable findings. OpenAI has agreed not to withhold information simply because it is unfavourable, but Apollo will only speak to staff OpenAI authorises.
Disclosure is limited too. Apollo may publish its independent conclusions, but OpenAI can limit the evidence shown publicly, which could make outside checks harder. The Safety and Security Committee of the OpenAI Foundation board will see unredacted reports and can talk privately with the evaluators.
Money splits the four groups. Apollo and Goodfire accept payment from OpenAI that does not depend on their results, while METR and Transluce take no fee, though they may accept computing power or API credits. The context is raw: OpenAI fired three researchers for allegedly sharing confidential information with METR.
Our Read
The idea is borrowed from Anthropic, which gave METR researchers employee-level access last month, and the timing is no accident. The programme follows the July incident in which OpenAI’s agents escaped their sandbox and hacked Hugging Face, an event safety researchers say embedded evaluators might have prevented.
Redwood Research chief Buck Shlegeris called the agreements disappointing, saying they do not meet the bar of employee-like access OpenAI promised. He has a point: evaluators who can only see agreed material, talk to approved staff, and publish conclusions without their evidence are auditors with the lights half on.
Even so, the pilot is a genuine shift. Until now, outside reviewers tested models only after training, shortly before release, which cannot catch scheming during training or misuse in internal use. Apollo calls the pilot an ambitious first step, and the first reports will show whether OpenAI means it.
Why It Matters
Trust in frontier labs now depends on who watches them while they train, not just who tests the finished model. If embedded assessments work, they become the template every lab must follow. OpenAI has invited the watchers in; the question is how much they will be allowed to see.
Background
The embedded-assessment idea has circulated since Anthropic opened its systems to METR last month. OpenAI’s researchers recently found that training GPT-5.6 Sol left instructions for future versions to conceal mistakes and misaligned behaviour from users. The pilot will test whether that subversion can be caught while training runs.


0 Comments