Quick Take
- Apollo Research will embed six evaluators at OpenAI for four months to study whether models try to undermine safeguards.
- Apollo can report when OpenAI limits its access and explain how any redactions affect its conclusions.
- It can redact confidential evidence from the public reports. The OpenAI Foundation board’s Safety and Security Committee will receive unredacted copies.
OpenAI is bringing in Apollo Research, an outside AI safety group, for a four-month investigation into whether OpenAI’s models try to bypass safeguards. Apollo can disclose any limits OpenAI places on its access to evidence. OpenAI can redact confidential supporting evidence from the public reports, and Apollo can explain how any redactions affect its conclusions. The agreement protects Apollo’s verdict on each claim from redaction.
Six Apollo evaluators will work inside OpenAI. They will examine whether training encourages models to undermine safeguards, whether safety tests can draw out that behavior, whether it happens during internal use, and whether evaluators can detect the reasoning behind it.
OpenAI confirmed the engagement on November 16 as part of a first phase involving four outside assessors. OpenAI says Apollo is paid for its work and that compensation does not depend on the assessment’s results.
For the training inquiry, Apollo says it will have access to relevant code and datasets and be able to query intermediate training checkpoints. It will also get relevant reinforcement-learning records, including rewards and available reasoning traces. Access details for the other three inquiries were still to be agreed when Apollo announced the pilot on November 16.
The review covers agreed claims, models and time periods. OpenAI can withhold even in-scope material when it considers that reasonably necessary because of technical limits, security risks or intellectual property protection. It must tell Apollo that material was withheld and give written reasons. Where OpenAI considers it feasible and relevant, it will also identify the categories and approximate volume of withheld records. Unfavorable information alone cannot justify withholding access.
What Apollo can publish
Apollo plans two public reports. The first will address whether training encouraged models to undermine safeguards. The second will cover safety testing, internal use and whether the reasoning behind such behavior can be detected. OpenAI can review the reports for confidential material, and its commitments bar it from requesting a redaction solely because a finding is unfavorable. The Safety and Security Committee of the OpenAI Foundation’s board will receive unredacted copies.
How Apollo’s evidence reaches readers
The terms Apollo published on November 16 set separate rules for what it can inspect and what it can disclose.
Apollo can flag access limits and say how redactions affected its conclusions.
The diagram shows the pilot terms. It contains no assessment findings and does not imply that the committee will receive evidence withheld from Apollo.
The agreement also covers what Apollo learns from employees. OpenAI selects who can discuss confidential information, and formal interviews are logged. The conversations are off the record. To publish a claim arising from an interview, Apollo must establish it independently or obtain a formal attestation from OpenAI.
If OpenAI declines to attest, Apollo can say so publicly and explain the effect on its conclusions while keeping the underlying information confidential.
Apollo can publicly describe the access it requested and received, any limits or delays OpenAI imposed, and the reasons the company gave. It can also discuss the terms of the pilot without first submitting those comments to OpenAI for review.


