Skip to main content

OpenAI signs agreements with four independent assessors

17 NOVEMBER 2026·2 MIN READ·5 SOURCES

OpenAI says it has signed agreements with four independent assessors to examine model alignment and safeguards through embedded assessments. The assessors are Apollo Research, METR, Transluce and Goodfire.

OpenAI signs agreements with four independent assessors

Key takeaways · 4

  • 01

    OpenAI signed agreements with Apollo Research, METR, Transluce and Goodfire.

  • 02

    The assessments will initially focus on alignment and safeguards during training and internal use, alongside a pilot review of OpenAI’s monitoring systems.

  • 03

    Apollo Research says its six-person team will work within OpenAI for four months.

  • 04

    Transluce says its six-week pilot will review internal alignment evaluations and will not provide employee-level access.

Access and initial scope

OpenAI describes embedded assessments as giving independent experts deep access to internal models and systems so they can evaluate alignment and safeguards over an extended period.[1] The listed access includes model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with researchers and engineers.[1] OpenAI says the initial assessments will focus on alignment and safeguards for models during training and internal use, complementing its safety cases.[1] The first phase also includes a pilot assessment of OpenAI’s internal monitoring systems.[1]

Four different assessment plans

Apollo Research will examine whether models might covertly act against developers’ intentions and whether safeguards would detect that behavior, according to OpenAI.[1] Apollo says its six-person team will work within OpenAI for four months across training, evaluations and internal deployment.[2] Its pilot will examine whether reasoning relevant to subversion would be detectable, but Apollo says the pilot does not yet fulfill the criteria it has set out.[2] Goodfire will embed interpretability researchers and use its platform to analyze model internals, OpenAI says.[1] Goodfire says its researchers will monitor training runs for known risks and emerging anomalous behavior.[3]

Evaluation coverage and access

OpenAI says METR’s first alignment assessment will examine misaligned behavior in its largest reinforcement-learning run to date, including failures that existing tests and monitors may have missed.[1] OpenAI says it intends for METR’s assessment to be broader, pending mutual agreement on scope.[1] Transluce will initially assess the quality and coverage of OpenAI’s post-training alignment evaluations, according to OpenAI.[1] Transluce says its six-week pilot will review whether internal evaluations capture behaviors such as deception, manipulation and reward hacking.[4] It says staff will have privileged access to a handful of models, including at least one internal-only model and raw chain-of-thought outputs, but not employee-level or employee-like access.[4]

Publication and independence terms

Apollo says it will publish two reports presenting its findings, methods and limitations.[2] It says OpenAI agreed not to withhold information or staff access, or request redactions, merely because findings are unfavorable; Apollo also says items may be withheld for technical infeasibility, security risks or intellectual-property protection.[2] OpenAI says Apollo Research and Goodfire accept compensation that is not contingent on results, while METR and Transluce may accept API credits or compute.[1] Transluce says it and its staff are not accepting payment for the pilot.[4] OpenAI says the effort supports commitments under the White House Accord on Super Intelligence and expects the first phase to lead to longer-term engagements.[1]

For teams building or evaluating models, embedded assessments offer a way to put external reviewers closer to internal development processes. The distinct scopes, access arrangements and publication terms are worth checking before treating findings as comparable or as a complete account of model risk.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 17 November 2026

    OpenAI signs agreements with four independent assessors

Sources

AI fluency, one session a day, built for your work.