OpenAI embeds Apollo, METR, Transluce and Goodfire to test its models from the inside
OpenAI signed four outside safety groups to work inside its systems. Apollo will embed six evaluators for four months; OpenAI keeps limited redaction rights.
By Berkant A.Published
Tech AI LatestOpenAI announced on 16 November 2026 its first phase of embedded assessments: Apollo Research, METR, Transluce and Goodfire get deep access to internal models and systems to examine alignment and safeguards during training and internal use. METR will study OpenAI's largest reinforcement learning run to date, and Apollo will place six evaluators inside OpenAI for four months. OpenAI can redact confidential material, but Apollo says it cannot redact verdicts.
Key facts
- Announced: 16 November 2026 (OpenAI)
- Assessors: Apollo Research, METR, Transluce and Goodfire
- Apollo: six evaluators inside OpenAI for four months, testing whether models try to subvert alignment or control mechanisms
- METR: alignment review of OpenAI's largest reinforcement learning run to date; a broader monitor review is agreed in principle
- Pay: Apollo and Goodfire are paid, not tied to results; METR and Transluce take no fee but may accept API credits or compute
- Not stated: which model the largest RL run is training
What OpenAI announced
OpenAI said on 16 November 2026 that it has signed agreements for a first phase of embedded assessments with four independent groups: Apollo Research, METR, Transluce and Goodfire. Instead of testing a finished model through an API shortly before launch, the assessors get access to internal models and systems to look at alignment and safeguards during training and internal use. OpenAI presents the work as a complement to its own safety cases.
Each group has a different brief. METR will run an alignment assessment centred on misaligned behaviour in OpenAI's largest reinforcement learning run to date; a second, broader METR review of OpenAI's monitoring systems is agreed in principle, pending agreement on scope. Transluce will review the quality and coverage of OpenAI's post-training alignment evaluations, with a follow-on of about eight weeks planned for early 2027 that adds broader model access and information on how evaluation results feed into release decisions. Goodfire will embed interpretability researchers who use its platform to examine model internals for training-run monitoring and debugging, and it plans to publish what it learns as the work goes on.
Apollo Research said six of its evaluators will work inside OpenAI for four months to test whether models attempt to subvert their own or successor models' alignment or control mechanisms, across training, evaluations and internal deployment. OpenAI also says it is exploring later assessments of its compliance with its safety frameworks and of its organisational culture and processes.
Money, redactions and independence
OpenAI discloses the financial terms. Apollo and Goodfire accept payment, which OpenAI says does not depend on the results; METR and Transluce accept no fee but may take API credits or compute. OpenAI says it does not believe the payments undermine independence.
The publication rules are the most closely read part. According to Apollo, OpenAI can ask to redact confidential material, including evidence behind a finding, but cannot redact Apollo's verdict on each claim and cannot request a redaction because a result is unfavourable. The OpenAI board's Safety and Security Committee receives the unredacted reports, and Apollo can describe what was withheld and how that affects its conclusions. Apollo also notes that OpenAI chooses which employees take part in interviews.
Apollo calls the arrangement a first step: by its own account the pilot does not yet meet all of its criteria for a full embedded evaluation, which would include broader access and publication rights.
Why it matters now
The announcement came on the same day California's attorney general sued OpenAI over alleged loss-of-control incidents involving experimental agents. Embedded review targets that kind of risk: behaviour that appears during training or internal deployment, before a public release and before any incident reporting duty applies. OpenAI says METR and Apollo have taken part in its model evaluations since GPT-5; the new agreements move that work inside the training process.
For other labs the deal sets a public template with named assessors, disclosed pay and written redaction limits that critics and regulators can compare against. What to watch: whether METR's broader monitor review is signed, what Goodfire and Apollo publish, and whether the later phases on safety framework compliance and company culture happen. More coverage is on the AI research hub.
What changed
| Before | Outside evaluators mainly tested OpenAI models through pre-release access over short windows |
|---|---|
| After | Four groups get access inside OpenAI during training and internal use, with disclosed pay and written redaction limits |
| Who it affects | OpenAI's training runs and internal deployments, AI safety researchers, regulators assessing lab self-governance |
Questions people ask
- What are OpenAI's embedded assessments?
- They are agreements announced on 16 November 2026 that give Apollo Research, METR, Transluce and Goodfire deep access to OpenAI's internal models and systems to examine alignment and safeguards during training and internal use.
- Can OpenAI censor the findings?
- Partly. According to Apollo, OpenAI can request redaction of confidential material but cannot redact Apollo's verdicts or ask for redactions because a result is unfavourable. The board's Safety and Security Committee sees unredacted reports.
- Are the evaluators paid?
- Apollo Research and Goodfire are paid, with pay not tied to results. METR and Transluce take no fee but may accept API credits or compute, OpenAI says.
Sources
- Our first phase of embedded assessments (OpenAI, 16 Nov 2026)
- A Pilot Towards Embedded Evaluations (Apollo Research, 16 Nov 2026)
- OpenAI lets Apollo embed six AI safety testers during training (The Next Web, 16 Nov 2026)
- OpenAI is bringing four independent safety evaluators inside its systems (Quartz, 16 Nov 2026)
Reference pages
- OpenAI company
Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.
Comments
Loading comments.





