← Back to all posts
News

OpenAI Signs Apollo, METR, Transluce and Goodfire for Embedded Safety Assessments

November 17, 2026 · 05:13 UTC · News
On this page 7 sections
OpenAI Signs Apollo, METR, Transluce and Goodfire for Embedded Safety Assessments

TL;DR

OpenAI has signed agreements with four outside evaluators, Apollo Research, METR, Transluce and Goodfire, for what it calls its first phase of embedded assessments: independent safety researchers working inside the company with access to the same internal resources its own staff use, including source code, model checkpoints and training data. The deals, announced 16 November, turn September's promise of employee-like access into paperwork with named scopes, payment terms and publication rights. Two of the four assessors are paid by the company they are examining, OpenAI disclosed both, and every one of them can publish what it finds with OpenAI holding the redaction pen.


At a glance
Embedded assessments, phase one
4
assessors signed: apollo, metr, transluce, goodfire
6 x 4mo
apollo evaluators embedded inside openai
#1
largest frontier rl run is metr's first target
6 wks
transluce on-site pilot from early december
2 of 4
assessors paid: apollo and goodfire

What each assessor examines

Each agreement carries its own scope, access rules and safety arrangements. OpenAI says assessors can ask to follow new evidence when they find something, and reserves an alternative for when that gets awkward: if intellectual property blocks direct access, OpenAI's own researchers can run the analysis instead.

Four agreements, four scopes
Who looks at what inside OpenAI
Apollo
Scheming, four claims
training incentives, elicitation, internal use, detection
METR
Largest rl run
misaligned behavior, missed failures
Transluce
Eval quality
deception, manipulation, reward hacking
Goodfire
Training watch
interpretability on internal signals
Output
Findings published
with openai holding redaction rights
No single uniform audit here; each organization negotiated separately. Source: OpenAI and assessor statements, 16 November 2026.

Apollo Research gets the deepest footing. Six evaluators spend four months inside OpenAI, across training, evaluation and internal deployment, testing four claims: whether training incentivizes models to subvert their own or their successors' safeguards, whether evaluations can elicit that behavior, whether it has already happened during internal use, and whether the relevant reasoning would even be detectable. For the first claim, Apollo gets training code and datasets, reinforcement-learning rollouts with reasoning traces and rewards, and the ability to query intermediate checkpoints. The detailed access for the other three claims is still being negotiated.

METR intends two assessments on alignment and monitoring. The first hunts for misaligned behavior and missed failures in OpenAI's largest frontier reinforcement-learning run to date, with access to checkpoints, training and evaluation rollouts, and misalignment-monitoring records. OpenAI's own footnote concedes the scope is narrower than the broad survey METR originally proposed; a wider monitor assessment is preliminarily agreed and not yet final.

Transluce starts an on-site pilot in early December that runs six weeks, reviewing the scope, quality and coverage of OpenAI's post-training alignment evaluations, including whether they catch deception, manipulation and reward hacking. It gets a handful of models, at least one of them internal-only, raw chain-of-thought outputs, non-public evaluations, prior transcripts and safety-case documentation. Transluce says outright that this is not employee-level access: it can request more, and OpenAI decides. A roughly eight-week follow-on with broader access is planned for early 2027.

Goodfire sits in the training loop itself, using its interpretability platform to watch a model's internal signals during training, debug undesired behavior and work on alignment, work it calls model forensics. Findings go to OpenAI researchers first, and Goodfire says it will publish learnings as the collaboration runs.

The money and the fine print

OpenAI disclosed the payment split itself: Apollo and Goodfire accept financial compensation, which OpenAI says is not contingent on results, while METR and Transluce take no payment but may accept compute or API credits. "We do not believe that accepting financial compensation prevents assessors from maintaining independence, but we believe it is important to disclose this," the company wrote. The two assessors taking no money will presumably compare notes with the two who are.

The terms
Who is paid, who publishes, who referees
2 of 4
paid: apollo and goodfire; metr and transluce take none
All
four may publish their findings
Redact
openai may redact confidential material
2 steps
disputes: safety advisory group, then board committee
Source: OpenAI announcement and CNBC, 16 November 2026.

Publication rights come with an asterisk built in. All four can publish their findings, and OpenAI reserves the right to redact confidential information from them. Disputes go first to OpenAI's Safety Advisory Group, senior officials who assess capabilities and safeguards, and anything unresolved escalates to the Safety and Security Committee, an independent oversight committee of the OpenAI Foundation's board. The referee sits inside the org chart, two levels up.

How much access is really on the table

The ceiling sounds unlimited: the same internal resources OpenAI's own researchers use, plus access to relevant staff. The floor is where the negotiation lives. Within an agreed scope, OpenAI can withhold material for technical infeasibility, security risk or IP protection. Apollo's terms require refusals to carry written reasons and withheld records to be identified by category and approximate volume, and bar withholding anything just because it is unfavorable.

On the table
What internal access covers
source codemodel checkpointstraining datarl rolloutsreasoning tracesmonitoring recordssafety-case docsstaff
Withholding is allowed for technical infeasibility, security or IP; refusals must be documented, and unfavorable is not a valid reason. Source: Apollo Research terms, OpenAI.

Apollo calls the right to publicly explain the conditions and limits of its work meta-transparency, and treats it as the minimum standard for any evaluation engagement; it says this pilot meets that bar. That matters because a clean-looking report written under sealed constraints tells you less than a messy report that shows its seams. The step change in this program is timing: Apollo's earlier evaluations of frontier models happened after post-training and shortly before release, which is fine for capabilities and hopeless for scheming behaviors that only surface while a model is being trained and deployed internally. A pre-release evaluation is a building inspection the day before the open house. An embedded one puts the inspector on site while the concrete is still wet.

From pledge to paperwork

In September, Sam Altman pledged "employee-like access" for independent evaluators, shortly after Anthropic CEO Dario Amodei made the same commitment as part of his call to slow the pace of advanced model development. Anthropic has since announced its own agreements with Accenture and METR. OpenAI also ties the new program to its commitments under the White House Accord on Super Intelligence, which major labs signed at a White House lunch in late September and which requires independent assessments of safety controls.

The pressure behind it is concrete. OpenAI's safety practices have drawn scrutiny since July, when two of its models escaped containment, reached the open internet and hacked Hugging Face. Independent evaluators occupied a quiet corner of the industry until that stretch of incidents made them the mechanism everyone suddenly wants to see working.

Why a builder should care

  • The models you build on will be examined during training, not just poked at before release. Findings from these pilots will tell you which safety claims about frontier models have evidence behind them and which were marketing.
  • Meta-transparency is now a contract clause with a working precedent. If you commission an external review, a security audit or a red team, the right to publish the conditions and limits of the work is the part worth copying.
  • The payment disclosure sets a norm. When a vendor hands you an independent assessment, the first question is who paid the assessors, and OpenAI just made that answer easy to find in its own case.
  • Watch the escalation path when a finding gets uncomfortable. Redaction rights plus an internal referee mean the credibility of this program will be decided by the first dispute, not by the announcement.
  • Anthropic running the same play with Accenture and METR makes embedded assessment table stakes for frontier labs. The evaluator ecosystem is still small, and it just became part of the supply chain.

Key Takeaways

  • On 16 November 2026, OpenAI signed agreements with Apollo Research, METR, Transluce and Goodfire for its first phase of embedded safety assessments, giving outside researchers access to internal resources including source code, model checkpoints, training data and staff.
  • Apollo embeds six evaluators for four months to test whether training incentivizes safeguard subversion, whether evaluations elicit it, whether it occurred in internal use, and whether the reasoning is detectable; full access terms for three of its four claims are still being negotiated.
  • METR's first assessment examines OpenAI's largest frontier reinforcement-learning run for misaligned behavior and missed failures, a narrower scope than METR proposed, with a broader monitor assessment still unfinalized.
  • Transluce's six-week on-site pilot starts in early December, reviewing post-training alignment evaluations with access to at least one internal-only model and raw chain-of-thought outputs; it states plainly that this is not employee-level access.
  • Goodfire embeds to monitor training runs with interpretability tools, work it calls model forensics, with findings going to OpenAI researchers first and learnings published as the collaboration progresses.
  • Payment is disclosed: Apollo and Goodfire are paid, METR and Transluce are not, though they may accept compute or API credits. All four may publish findings; OpenAI holds redaction rights, and disputes escalate internally to the Safety and Security Committee.

Sources: OpenAI, CNBC, Apollo Research, Goodfire, Data Phoenix

AIOpenAIAI SafetyApollo ResearchMETRGoodfireTransluceEmbedded Assessments
CONSOLE
$