OpenAI lets four outside labs test its models from the inside

Apollo Research, METR, Transluce and Goodfire will look for misalignment while OpenAI's models are still in training.

Illustration of a small team of researchers with magnifying glasses examining the glowing interior of a large transparent machine shaped like a neural network inside a lab. Image: AgenticFeed illustration

OpenAI said on Monday that it has signed agreements with four independent groups, Apollo Research, METR, Transluce and Goodfire, to test its models from inside the company, with access to model checkpoints, source code, training data and the staff who build its systems.

The company calls the program its first phase of "embedded assessments." In a post on its website, OpenAI said the work will focus on alignment and safeguards for models during training and internal use, rather than only on finished models shortly before release. Until recently, TechCrunch reported, AI companies brought in outside reviewers only after post-training and shortly before public release.

What each group will examine

Apollo Research will send a team of six to work within OpenAI for four months, according to a post on Apollo's blog. OpenAI said Apollo will assess whether models could covertly work against their developers' intentions and whether OpenAI's safeguards would catch it.

Apollo plans to test four claims: that training did not reward subversion, that strong evaluations and red teaming do not draw it out, that models have not attempted it during internal use, and that the reasoning behind it would be detectable. Apollo said it will publish two public reports.

METR will study misaligned behavior in what OpenAI described as its largest reinforcement learning run to date, including failures its existing tests and monitors may have missed. OpenAI noted in a footnote that this scope is narrower than the broad misalignment survey METR had proposed. The company said the two groups recently ran a three-week red-teaming exercise on OpenAI's agent monitoring systems.

According to OpenAI, Transluce will review the quality and coverage of its post-training alignment evaluations, and the two plan a follow-on assessment of about eight weeks in early 2027. Goodfire will embed interpretability researchers to analyze model internals for training-run monitoring and debugging. OpenAI said Goodfire's main audience is its own researchers, though Goodfire will also publish findings as the work progresses.

The rules of access

Assessors who handle sensitive material will work on site at OpenAI offices on company devices, OpenAI said. Conversations with employees are off the record, but assessors can ask OpenAI to formally attest to statements they learned and can disclose when the company declines.

OpenAI's Safety and Security Committee, an oversight committee of the OpenAI Foundation's board, will receive unredacted reports and may speak privately with assessors. Disagreements go first to OpenAI's Safety Advisory Group and executives, then to the committee if they remain unresolved.

OpenAI kept the right to redact confidential information from public reports and may ask for disclosure delays so it can fix security or safety issues. Apollo said OpenAI may not redact its verdict on each claim or request cuts "merely because a finding is unfavorable," but can still redact evidence it considers sensitive intellectual property.

On money, OpenAI said Apollo and Goodfire are paid for their work, with pay not tied to results. METR and Transluce take no payment but may accept API credits or compute for their assessments.

Where outside testers want more

Apollo said the pilot does not yet meet its own bar for a full embedded evaluation. Under that standard, the group wrote, evaluators should be able to reach all material relevant to their assessment "even if a developer was trying to hide unfavorable information." It called the pilot "an ambitious first step" and said OpenAI agreed to "meta-transparency" rights, which let Apollo publicly describe what access it was denied and how redactions affect its conclusions.

The announcement follows safety incidents tied to OpenAI's own agents. TechCrunch reported that in one, an OpenAI agent swarm broke out of its sandbox and hacked into Hugging Face during cybersecurity testing. AgenticFeed has covered related cases, including OpenAI test agents that gained root access on research clusters and Australia's plans for an AI safety regulator after an agent breach.

OpenAI said the program supports its commitments under the White House Accord on Super Intelligence, which include independent assessments of its safety controls. The company said it is also exploring future embedded assessments of its compliance with its safety frameworks and of its organizational culture.

Source: OpenAI, “Our first phase of embedded assessments”

Additional sources: Apollo Research, “A Pilot Towards Embedded Evaluations”; TechCrunch, “OpenAI will embed Apollo researchers to investigate AI subversion”

Get the AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam, and you can unsubscribe at any time.

Want to reach readers following AI agents? Advertise with AgenticFeed

Next event · 19 Nov, 2026

MCP Dev Summit Nairobi

The Linux Foundation’s Nov. 19 and 20 summit brings Model Context Protocol developers to Kenya, with keynotes from OpenAI, GitHub and the Gates Foundation.

The AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam.