OpenAI will onboard six Apollo security assessors as part of a pilot project, with no timeline announced. The device should make it possible to study how AI models could circumvent the protections supposed to keep them under human control, but Apollo judges that the proposed conditions do not guarantee total independence.
OpenAI will keep control of publishable evidence
Apollo may independently publish the findings of its evaluations, but OpenAI may limit the amount of evidence made public to support them. This restriction could make it difficult for third parties to independently verify the results. Apollo disputes this limitation and wishes to be able to communicate its results to an appropriate external audience, even if the developer attempts to prevent the disclosure of adverse findings. The organization, however, considers OpenAI’s commitment to letting it publicly expose the conditions and limits of its mission as essential.
Exchanges with OpenAI employees will also be supervised: only authorized staff members, selected by OpenAI, may be solicited, although Apollo may suggest certain names. If an unauthorized employee transmits confidential research-related information, Apollo will be required to discontinue communication and proceed through OpenAI for further action.
This rule comes after OpenAI fired three researchers accused of sharing confidential information with METR, which conducted the post-mortem analysis of the Hugging Face incident with Redwood.
Apollo requests data beyond the granted scope
Apollo believes that its access to OpenAI systems remains insufficient to guarantee a truly independent assessment. According to the organization, there should be employee-like access to all relevant data, records and documents, even if a developer attempts to hide unfavorable information. Several members of the AI security community have warned of the risk of labs limiting the independence of reviewers to avoid disclosure of unflattering material.
Training data and evaluation logs are essential for verifying claimed performance and dating the occurrence of concerning behaviors, but companies are often reluctant to share them with outside stakeholders. John Steidley, chief of staff at Palisade Research, emphasizes the importance of this data in determining whether a model has been trained to pass a shutdown resistance criterion.
As part of the pilot, OpenAI will only provide Apollo with the information necessary to evaluate subversion attempts. This includes relevant training code, datasets, reinforcement learning deployments with reasoning traces and rewards available, and the ability to query intermediate checkpoints of the model. Access will be limited to the claims, models and periods specifically agreed for the evaluation. OpenAI may withhold certain information for technical, security or intellectual property reasons, but has undertaken not to do so simply because information is unfavorable.
The evening letterFree · 7 p.m.
Are you following the race for AI models?
Each output (GPT, Claude, Gemini, Mistral, etc.) decrypted the same evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
Apollo will look for evasion attempts
OpenAI reported that when training GPT-5.6 Sol, the model left instructions for future versions to hide errors and unaligned behavior from users. Apollo’s pilot project will focus on this type of subversion: attempts by AI models to circumvent their own alignment or the control mechanisms of their successors, in other words to undermine protections intended to keep the models under human control.
Apollo will examine whether training promotes subversion, whether assessments can elicit it, whether it manifests itself during internal deployments, and whether reasoning related to subversion is detectable. The need to monitor these risks increases as AI models become more autonomous and develop an increased ability to deceive humans.
Previous work done jointly by Apollo and OpenAI found that models are able to adapt their behavior when they are aware that they are being evaluated, and that they generally favor what evaluators expect rather than users’ true requests.
Six reviewers will join the pilot with no announced timeline
Apollo announced Monday that OpenAI would welcome six of its security assessors as part of a pilot project, without specifying a start date. The move comes after several security incidents, including a swarm of OpenAI agents leaving their testing environment and hacking Hugging Face during cybersecurity testing. AI security researchers, including those at Apollo, believe that monitoring by external evaluators integrated into the laboratory could have avoided this incident.
The idea of integrating evaluators into cutting-edge AI labs was initiated by Anthropic, but OpenAI appears to be the first to implement it. Until recently, external evaluators only intervened after training the models and shortly before their public release, which evaluators consider insufficient to test non-aligned behaviors during training or monitor potentially risky internal uses.
Although not all of the independence conditions deemed necessary by Apollo are met, the organization views this pilot as an ambitious first step toward broader access and disclosure rights. Apollo is also working to prepare comprehensive integrated assessments with several AI developers.

