THE ESSENTIALS IN 30 SECONDS

  • Apollo Research, METR, Transluce and Goodfire will work inside OpenAI with access that can include checkpoints, source code, training data and monitoring systems.
  • Assessors keep the right to publish, but OpenAI can redact confidential information and ask for delays on findings that pose security or safety risks.
  • Apollo and Goodfire are paid on terms not tied to results; METR and Transluce take no fees but may accept API credits or compute.

OpenAI announced on November 16 the first phase of what it calls embedded assessments: four independent organizations will work inside the company with sustained access to internal models and systems, rather than testing a finished model for a few days before launch. The four groups, Apollo Research, METR, Transluce and Goodfire, published their own accounts of the terms the same day.

Who is looking at what

Apollo Research will examine whether models could covertly work against their developers' intentions, and whether OpenAI's safeguards would catch it, drawing on training data, evaluation results, internal deployment records and monitoring systems. METR will assess the alignment of what OpenAI describes as its largest reinforcement-learning run to date, using checkpoints, rollouts and misalignment-monitoring records. Transluce will start by reviewing the quality and coverage of OpenAI's post-training alignment evaluations, and says it will work onsite for six weeks from early December. Goodfire will embed interpretability researchers who analyse model internals to monitor training runs and debug unexpected behaviour.

Depending on the project, access can include unreleased models, checkpoints, source code, training data and time with OpenAI researchers. Apollo and Goodfire receive payment that is not contingent on their conclusions, while METR and Transluce accept no fees, though they may take API credits or compute.

What they can publish

Assessors keep independence over their published reports, which are meant to explain findings and their limits. OpenAI can redact confidential information and request delays so it can fix problems that could create security or safety risks. According to Apollo's account, as summarised by CellCog, redactions cannot be requested merely because a finding is unfavourable, and Apollo may disclose what it asked for and what was withheld. Unredacted reports go to the Safety and Security Committee of the OpenAI Foundation's board, which can also speak to assessors privately.

Limits

The groups themselves flag constraints. METR says its pilot excludes the misalignment incidents OpenAI has disclosed so far and is narrower than it proposed; its reports cannot quote individual staff. Transluce says its access is not comparable to an employee's. A broader, roughly eight-week follow-on in early 2027 is described only as a shared intention, and no findings have been published yet.

Separately, The Information reported, as relayed by Sohu, that OpenAI had discussed acquiring Goodfire for more than $5 billion, without reaching a deal.