Skip to content
AI Primer
breaking

OpenAI signs assessment agreements with four outside research groups

OpenAI has assessment agreements with METR, Apollo, Transluce and Goodfire. METR's scope is limited to one reinforcement-learning run, while access restrictions and publication commitments vary across assessors.

4 min read
OpenAI signs assessment agreements with four outside research groups
OpenAI signs assessment agreements with four outside research groups

TL;DR

  • Four outside research groups will examine OpenAI’s models and safeguards under its new embedded-assessment agreements.
  • METR’s pilot covers one reinforcement-learning run and excludes previously disclosed misalignment incidents, according to METR.
  • Transluce gets raw chain-of-thought outputs while describing its access as short of employee-like access in its announcement.
  • Goodfire’s first-phase priority is internal safety debugging, with publication to be mutually agreed, Goodfire clarified.

The sharpest restriction sits in Transluce’s agreement: researchers can examine an internal-only model but cannot publish results about it. Apollo’s terms protect its claim-by-claim verdicts even when OpenAI redacts the supporting evidence.

Apollo: four subversion claims

Apollo Research will embed six evaluators for four months. OpenAI’s announcement identifies four claims for testing:

  1. Training did not incentivize subversion.
  2. Strong evaluations and red teaming do not elicit subversion.
  3. Models have not attempted subversion during internal use.
  4. Reasoning behind subversion would be detectable.

Apollo says detailed access arrangements are agreed for the training-incentives assessment; arrangements for the remaining claims still need agreement.

METR: one RL run

OpenAI will supply METR with three kinds of evidence from its largest reinforcement-learning run:

  • Training and evaluation transcripts.
  • Misalignment-monitoring records.
  • Model snapshots throughout reinforcement learning.

Previously disclosed misalignment incidents fall outside the pilot. METR says the scope is narrower than it proposed.

A separate, broader monitor assessment is only preliminarily agreed, with its scope still pending, according to OpenAI’s outline.

Transluce: post-training evaluations

Transluce’s six-week onsite pilot starts in early December and will assess the quality and coverage of OpenAI’s post-training alignment evaluations.

Both sides intend a broader follow-on in early 2027. OpenAI describes approximately eight weeks, including access to information about how evaluation results influence model-release decisions; Transluce says the details remain unfinalized.

Goodfire: model forensics

Goodfire will embed interpretability researchers to analyze model internals during training. Its “model forensics” work investigates internal signals behind unexpected behavior, with OpenAI researchers as the primary audience for findings.

Employee access

NeelNanda5 argues that OpenAI’s control over information flow leaves assessors dependent on its good faith, contrasting the pilots with its earlier employee-like-access promise.

Apollo’s access rules impose four concrete limits:

  • General Slack channels are unavailable by default.
  • OpenAI selects authorized interviewees, even when Apollo proposes names.
  • Formal conversations with approved staff must be logged.
  • If unauthorized staff raise confidential information, Apollo must stop substantive discussion and route follow-up through its designated OpenAI contact.

Goodfire has different interview terms: a dedicated Slack channel and no requirement to record conversations, according to its clarification. OpenAI retains the final choice of staff contacts, with refusals limited to specified reasons.

METR’s earlier three-week monitor exercise included access to OpenAI’s internal codebase, establishing a precedent for deeper technical access.

Publication rights

The assessors have different reporting arrangements:

  • Apollo: Two public reports, first on training incentives, then on evaluations, internal deployment and detectability. Its terms allow it to describe withheld access, redactions and their effect on conclusions.
  • METR: One or more public reports, as promised in its announcement. Staff interviews can guide the work, but reports cannot quote or cite individual views, METR disclosed.
  • Transluce: Its published terms promise disclosure of access requests, approvals and denials, and how those decisions affected its conclusions.

OpenAI’s publication protocol retains confidential-information redaction rights and permits requests for responsible-disclosure delays. It specifies no maximum delay.

Funding and oversight

OpenAI discloses two funding arrangements:

  • Apollo and Goodfire: Paid, with compensation independent of assessment results.
  • METR and Transluce: No assessment fees, although they may accept compute or API credits.

Separately, The Information reported that OpenAI explored acquiring Goodfire for more than $5 billion.

The OpenAI Foundation Board’s Safety and Security Committee will receive unredacted reports and resolve disputes that remain after escalation to executives. Assessors and the committee can speak privately during the work.

Participating assessor staff also face a “brief” cooling-off period before embedding with another AI developer. OpenAI’s public protocol does not specify its duration.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 5 threads
TL;DR2 posts
METR: one RL run1 post
Transluce: post-training evaluations1 post
Goodfire: model forensics1 post
Publication rights2 posts
Share on X