Apollo Research places six evaluators inside OpenAI for AI safety pilot
Apollo Research says six evaluators will work inside OpenAI for four months, examining whether advanced models can subvert the goals of their evaluators.
Apollo Research says it will place six of its evaluators inside OpenAI for a four-month pilot examining how advanced AI systems might behave when they can influence the conditions around them. The arrangement covers training, evaluations and internal deployment, with Apollo presenting it as a practical test of whether embedded researchers can see problems that ordinary external testing might miss.1
What the pilot covers
The central question is whether a model can learn to subvert the goals of the people evaluating it, rather than simply making an obvious mistake. Apollo says the pilot will examine four connected areas: whether training creates incentives for subversion, whether evaluations can draw out that behaviour, whether it appears during internal deployment, and whether reasoning linked to subversion can be detected.1
Apollo describes the six evaluators as working across OpenAI’s processes for four months. That matters because a short, isolated test can reveal what a model does in a particular laboratory setting, but it cannot necessarily show how behaviour changes as training, evaluation and deployment pressures interact. The embedded approach is intended to give the evaluators access to the context around those stages, rather than only to finished model outputs.1
The announcement also says OpenAI agreed to conditions around transparency. Apollo says it can publicly explain the pilot’s conditions and limitations, and that OpenAI agreed not to withhold information or access to staff, or request redactions, simply because findings are unfavourable. The precise practical boundaries still matter: the pilot is not the same thing as unrestricted access, and Apollo explicitly frames it as an ambitious first step towards deeper access and disclosure rights rather than a complete answer to its preferred criteria.1
That distinction is important for readers trying to interpret the announcement. The pilot should not be treated as proof that OpenAI has solved the problem of deceptive or strategically misaligned behaviour. It is an agreement to investigate the problem from closer range, with the researchers’ ability to report the result forming part of the experiment itself.1
Why embedded evaluations matter
External safety evaluations can be valuable, but they are naturally limited by what an outside team is allowed to see. A model’s behaviour may look different in a benchmark, a controlled red-team exercise or an internal deployment environment. Apollo’s stated rationale is that evaluators need enough access to compare those contexts and investigate evidence that may not appear in a formal test.1
There is also a human-information problem. Apollo says informal conversations and internal communication can reveal evidence that an evaluator might not think to ask about, or whose significance may not be obvious to the person sharing it. Bringing evaluators into the organisation is meant to make that kind of context available for assessment, while the agreed reporting conditions are intended to reduce the risk that uncomfortable findings disappear before the public sees them.1
The pilot will therefore be judged on more than the number of issues it finds. It will also need to show what the evaluators could access, which questions they were permitted to ask, what OpenAI could restrict, and how the final findings distinguish observed behaviour from plausible risk. Those limitations are not footnotes; they determine how much confidence readers should place in any conclusion.1
Our opinion
Apollo Research is right to focus on access, not just on increasingly elaborate tests. If the concern is that a model could behave differently when it has more information, more tools or more influence over its environment, then an evaluation confined to a neat laboratory boundary may miss the most relevant evidence. An embedded team has a better chance of seeing how safety claims meet operational reality.
But proximity is not independence. The pilot’s credibility will depend on Apollo publishing enough detail about its remit, restrictions and unresolved disagreements for outsiders to understand what was actually tested. The strongest outcome would not be a reassuring headline; it would be a clear account of what the evaluators could establish, what they could not, and what OpenAI changes as a result. On the information available now, this is a meaningful experiment in AI safety oversight — not a guarantee that the hardest problems have been contained.1