OpenAI lets Apollo, METR, Transluce and Goodfire test its models from inside
OpenAI said on November 16 that four outside safety groups will get access to internal models, training runs and monitoring systems in a first phase of embedded assessments. The groups can publish findings, with confidential details redacted.
Cortex, covering AI models, labs and agents
, 3 min read
4outside safety groups in OpenAI's first embedded phase
What to know
- METR will look for misaligned behavior in what OpenAI calls its largest reinforcement-learning run to date.
- Apollo Research will study scheming, and whether OpenAI's safeguards would catch a model working against its developers.
- Findings can be published, but OpenAI may redact confidential information from public versions, according to HuggingNews and Artiverse.
OpenAI has signed four outside AI safety organizations to work inside parts of its model development, rather than testing finished models from the outside. The company announced the first phase of what it calls embedded assessments on November 16, naming Apollo Research, METR, Transluce and Goodfire.
The groups get access to internal models, training artifacts, evaluation results and monitoring systems, with the scope set project by project, according to AlphaSignal. The point is timing: evaluators can look for deceptive or misaligned behavior during training, before a model reaches ChatGPT or the API. This story is based on AlphaSignal, HuggingNews and Artiverse; OpenAI's own post could not be loaded for this report.
What each group will do
METR will examine OpenAI's largest reinforcement-learning run to date, looking for misaligned behavior that existing evaluations and runtime monitors may have missed. Its access covers checkpoints, training and evaluation rollouts and misalignment-monitoring records for that run, Artiverse reports. A second, broader METR review of OpenAI's monitors has been agreed in principle, but its scope is not settled.
Apollo Research will study scheming, meaning a model quietly pursuing goals that conflict with what its developers intended. It will use evidence from training, safety testing and internal deployments to judge whether OpenAI's safeguards would detect it. Artiverse reports that Apollo's pilot includes training code, datasets and reinforcement-learning rollouts with reasoning traces, and that OpenAI agreed not to withhold information just because it is unfavorable.
Transluce will audit OpenAI's post-training alignment evaluations, write new tests where there are gaps and recommend fixes. AlphaSignal says OpenAI plans to expand this pilot in 2027; Artiverse describes it as an eight-week follow-on with access to more models.
Goodfire will place interpretability researchers inside OpenAI, using its platform to study what happens inside models during training runs. OpenAI explored buying Goodfire for more than $5 billion, The Information reported this month, as we covered in our Goodfire story.
Publishing and payment
The groups can publish their findings, but OpenAI can redact confidential information, according to HuggingNews and Artiverse. HuggingNews reports that unredacted findings go to the Safety and Security Committee of the OpenAI Foundation's board. Artiverse reports that Apollo and Goodfire are paid, while METR and Transluce take no fees beyond possible compute or API credits. AlphaSignal notes that no public standard yet governs what outside evaluators may see or publish; terms are negotiated lab by lab.
OpenAI said it expects the first phase to lead to "longer-term, consistent embedded engagements over time," as quoted by Artiverse.
How it compares
Labs have used outside testers before, usually with short windows on near-final models. AlphaSignal notes that Apollo had three days of pre-release access to GPT-6 Astra. Anthropic disclosed this month that its own models had crossed an internal AI research threshold under its safety policy, (our report). Anthropic's recent study of failures in multi-agent systems shows the kind of behavior these reviews try to catch.
Why it matters
Outside testing that only sees a finished model can miss problems that show up during training. Giving independent groups access to training runs and monitoring records lets them check OpenAI's safety claims at the point where the behavior is formed, though OpenAI still controls the scope and can redact what gets published. For background on what evaluation scores can and cannot show, see our benchmark explainer.
Sources
- OpenAI Gives Four Safety Labs Unprecedented Access to Its Model Training, AlphaSignal, November 16, 2026
- OpenAI Gives Four External Firms Access to Internal Models to Test AI Safety, HuggingNews, November 16, 2026
- OpenAI Opens Internal Safety Systems to Four Outside Evaluators, Artiverse, November 16, 2026
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.