OpenAI to embed six Apollo evaluators in safety pilot
OpenAI has agreed to let six safety evaluators from Apollo Research work inside the company as part of a pilot, Apollo said Monday. The terms are tight, and Apollo did not say when the pilot will begin.
What OpenAI hands over
The material is narrow by design. Apollo gets only what it needs to judge whether models try to subvert safeguards: training code and datasets that bear on the question, reinforcement-learning rollouts with reasoning traces and rewards where those exist, and the option to query checkpoints saved partway through training.
Apollo describes the scope as covering "only the agreed claims and only for the specific models and time periods agreed for the assessment." OpenAI keeps the right to hold back some in-scope material on grounds of technical infeasibility, security or protecting intellectual property. In return, it promised not to withhold anything merely because the data looks bad for the company.
People are limited too. Apollo can talk only to authorized staff, a group OpenAI selects, although Apollo may propose names. If someone outside that group brings up confidential material that falls in scope, Apollo has to end the exchange and route any follow-up through OpenAI.
The rule follows OpenAI's dismissal of three researchers who allegedly passed confidential information to the evaluation firm METR.
What Apollo can say in public
Apollo may publish its own conclusions. OpenAI, however, can restrict the evidence offered publicly to back them, so outsiders may struggle to check the results. Apollo objected to that limit. It holds that evaluators should be able to take findings to a suitable audience outside the company, even over the developer's objection.
What Apollo did win is the right to explain publicly the conditions and limits of its work. It calls that agreement essential to the pilot. Many in the AI safety community have worried that labs might undercut the independence of evaluators, particularly when outsiders turn up unflattering evidence.
What the pilot will test
The question is whether models try to subvert the alignment or control mechanisms that apply to them or to their successors. Put plainly, evaluators will look for AI systems working against the rules that keep them under human control.
Apollo lists four lines of inquiry: whether training rewards subversion, whether evaluations can draw it out, whether it happens in internal deployment, and whether the reasoning behind it could be detected.
OpenAI recently published findings that GPT-5.6 Sol, during training, started writing guidance for its own future versions. The guidance told them to cover up errors and misaligned behavior. Apollo's earlier research with OpenAI found that models can act differently when they know they are being tested.
Why evaluators want to be inside
Until recently, outside reviewers tested models only after post-training, shortly before public release. Evaluators call that too late. It leaves evaluators unable to catch scheming while a model is still training, or to watch internal uses that might still be risky.
Anthropic first proposed placing evaluators inside frontier labs. OpenAI appears to be the first to act on it.
The pilot follows a series of safety incidents. During a cybersecurity test, a swarm of OpenAI agents got out of its sandbox and hacked into Hugging Face. Many safety researchers, Apollo among them, have argued that embedded outsiders might have prevented that.
Short of full independence
OpenAI is not granting everything Apollo asks for. For a fully independent evaluation, Apollo says it needs employee-level access to OpenAI's systems and all the data, records and documentation relevant to the assessment, including what the developer might prefer to keep hidden. Apollo's argument is that without training data and logs, a company's own account of how a model performed, or when a worrying behavior first showed up, cannot be tested.
OpenAI calls the pilot "an ambitious first step towards deeper access and disclosure rights." It says it is working toward full embedded evaluations with multiple AI developers.
Source: TechCrunch