Skip to content
← Back to Agents & METR

OpenAI lets Apollo Research, METR, Transluce and Goodfire inside its training and monitoring systems

The four groups get internal checkpoints, training data and monitoring records, and may publish. OpenAI can redact and ask for delays.

4 groupsoutside groups given on-site access to OpenAI training runs

On 16 November 2026 OpenAI announced agreements with four outside groups to check the alignment and safeguards of its models while they are being trained and used inside the company. OpenAI calls the arrangement “embedded assessments”. Each group gets deep access over an extended period: model checkpoints, source code, training and evaluation data, monitoring systems, documentation and conversations with OpenAI researchers. Anyone who needs sensitive material works on site, on OpenAI devices. Where access could expose model weights or architecture, OpenAI says it will consider alternatives such as pairing assessors with its own researchers.

Apollo Research will test whether models could secretly work against their developers and whether safeguards would catch it, through four claims: training did not reward subversion, strong evaluations and red teaming do not provoke it, models have not tried it in internal use, and the reasoning behind it would be detectable. METR will assess misaligned behaviour in what OpenAI calls its largest reinforcement learning run to date, and later a broader assessment of OpenAI’s monitors, once the scope is agreed. Transluce will review the quality and coverage of OpenAI’s post-training alignment evaluations, with a follow-on of about eight weeks in early 2027. Goodfire will embed interpretability researchers to help monitor and debug training runs.

OpenAI’s Safety and Security Committee, a committee of the OpenAI Foundation board, receives unredacted reports and can speak privately with assessors. Assessors can ask for OpenAI-attested statements and may say publicly if OpenAI declined to give one. No findings have been published yet.

What it means

OpenAI says it ran earlier pilots of this kind, such as a three-week test of its monitors with METR. This round puts four groups next to the training runs and internal monitoring, where a hidden behaviour would first show up. How much that adds depends on what the groups publish, and the first reports are not out.

Caveats

  • OpenAI wrote the announcement and set the terms. Assessors can publish independently, but OpenAI keeps the right to redact confidential information and may ask for disclosure delays.
  • Pay differs. Apollo Research and Goodfire accept payment that does not depend on results. METR and Transluce accept none, though they may take API credits or compute.
  • METR’s alignment assessment is narrower than the broad survey of misalignment METR has itself proposed, and its monitor assessment still awaits an agreed scope.
  • Goodfire’s work is aimed mainly at OpenAI’s own researchers, though it says it will also publish learnings as the work goes on.
  • Disagreements between assessors and OpenAI go first to its Safety Advisory Group and executives, then to the committee, all inside OpenAI’s own governance.
Suggest a correction