OpenAI’s embedded evaluator pilot places a team of six from Apollo Research inside the company for four months to test whether its models try to subvert their own safety controls. Under terms Apollo published on November 16, 2026, OpenAI cannot redact Apollo’s verdicts, and it can redact the evidence behind those verdicts when it considers that evidence sensitive intellectual property.
The OpenAI Embedded Evaluator Pilot Covers Training, Evaluations and Internal Deployment
Apollo will examine four questions across the development cycle: whether training gives models an incentive to subvert their alignment or control mechanisms, whether evaluations can draw that behavior out, whether it happens during internal deployment, and whether reasoning tied to it would be detectable. OpenAI announced on Monday that researchers from Apollo, Goodfire, METR and Transluce would join it in separate pilots, with Apollo starting this week, TechCrunch reported. OpenAI said the work supports its commitment under the White House Accord on Super Intelligence to rely on independent assessments of its safety controls. Apollo’s earlier evaluations of OpenAI models happened after post-training and shortly before public release. Apollo says that timing cannot catch scheming that emerges during training or misuse that happens only in internal deployment.
Apollo Gets Training Code and Checkpoints for Its First Question
For the question about training incentives, OpenAI committed to give Apollo the relevant training code and datasets, reinforcement learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints. Access for the other three questions has yet to be agreed. Inside the agreed scope, OpenAI may still withhold items it considers technically infeasible to share, a security risk or protected intellectual property. It agreed not to withhold anything because it is unfavorable, and it must tell Apollo when withheld items exist and give written reasons. All of Apollo’s work runs on OpenAI hardware, and by default the evaluators have no access to general internal channels such as Slack.
OpenAI Chooses Which Employees the Evaluators Can Talk To
Apollo may discuss confidential OpenAI information only with authorized staff, and OpenAI selects those people, though Apollo can request expertise and suggest names. If an employee outside that group raises confidential information that falls within the pilot’s scope, Apollo must stop the conversation and route any follow-up through its designated OpenAI contact. TechCrunch notes that this condition follows OpenAI’s firing of three researchers in October over the alleged sharing of confidential information with METR. Claims Apollo draws from employee conversations also need OpenAI’s formal attestation before they can be published, unless Apollo establishes them independently. Apollo can say publicly that an attestation was declined and explain how that limits its conclusions.
OpenAI Cannot Redact Apollo’s Verdicts and Can Redact Supporting Evidence
Apollo will publish two reports. The first covers whether training incentivizes subversion. The second covers evaluations, internal deployment and the detectability of subversive reasoning. OpenAI reviews both and may request redactions of confidential information, including evidence that supports Apollo’s findings. It may not redact the verdict on any claim, and it may not request a redaction because a finding is unfavorable. Readers may therefore see a conclusion without all of the evidence behind it. Apollo says it will be able to state whether redactions affected its conclusions. OpenAI will share the unredacted reports with the Safety and Security Committee of its board.
Meta-Transparency Is the Condition Apollo Treats as a Minimum
Apollo says the pilot does not yet meet its own criteria for full access and disclosure, and that it does meet its bar for meta-transparency, the right to explain the limits of its work in public. That right lets Apollo describe what it requested and received, what was withheld or delayed and why, what was removed from a report and how much it mattered, and any pressure it believed shaped its work, all without prior review by OpenAI. Apollo argues that without this right a clean report can look reassuring when it only reflects what an evaluator was allowed to examine. Buck Shlegeris, chief executive of Redwood Research, wrote on Monday that the agreements fall short of employee-like access and called them “disappointing but not surprising.” TechCrunch also reports that Apollo and Goodfire are paid for their work, while METR and Transluce are not paid and may accept compute or API credits.
Why This Matters for Engineering Leadership
Apollo’s published terms are a working template for judging any independent review, including the AI vendor assessments and model audits engineering leaders commission or rely on. The useful questions are the ones Apollo answered in public. Who chooses the people the reviewer can interview? What can the vendor withhold, and does it have to say so in writing? Can the reviewer publish its verdict without edits, and can it say how redactions changed that verdict? A reviewer that cannot describe the limits of its own access gives a buyer a clean report with no way to tell how much it covered. When your team signs an AI vendor contract or accepts a third-party safety attestation, ask for the equivalent of Apollo’s meta-transparency clause and read the scope section before the conclusions.
FAQ
What is OpenAI’s embedded evaluator pilot?
It is a set of pilots that place outside safety researchers from Apollo Research, Goodfire, METR and Transluce inside OpenAI. Apollo’s pilot puts six evaluators inside the company for four months.
What will Apollo Research test at OpenAI?
Apollo will test whether OpenAI’s models try to subvert their own or their successors’ alignment or control mechanisms. It will look at training incentives, evaluations, internal deployment and whether subversive reasoning is detectable.
Can OpenAI redact Apollo’s findings?
OpenAI cannot redact Apollo’s verdict on each claim or request redactions because a finding is unfavorable. It can redact supporting evidence it considers sensitive intellectual property.
Who decides which OpenAI employees Apollo can talk to?
OpenAI selects the authorized staff, though Apollo can suggest people. If another employee raises confidential in-scope information, Apollo must stop and route follow-up through its designated OpenAI contact.
Is Apollo Research paid by OpenAI for the pilot?
Yes. TechCrunch reports that Apollo and Goodfire receive financial compensation, while METR and Transluce do not but may accept compute or API credits.
What is meta-transparency in an AI safety evaluation?
It is the evaluator’s right to publicly explain the conditions and limits of its work. Apollo says OpenAI’s agreement meets its bar for meta-transparency.
