Apollo plans to examine OpenAI models from inside the lab
UK-based Apollo Research has agreed a pilot that puts six evaluators inside OpenAI for four months. They will investigate whether models try to undermine human control during development and internal use. The agreement sets boundaries on access to training information and staff, while preserving public explanations of those restrictions and plans for two reports.
Artificial Intelligence··Night
Evaluators follow development inside OpenAI
Apollo Research, a UK-based AI safety evaluation organisation, has agreed an embedded pilot with OpenAI to investigate whether models attempt to undermine the mechanisms keeping them under human control.[1], [2]
The proposed team has six evaluators working for four months. Its questions cover incentives created by training, behaviour elicited in evaluations, internal deployment and the visibility of relevant reasoning. Earlier Apollo evaluations generally came after post-training, shortly before public release.[1]
Training information opens under agreed conditions
Access for the training inquiry includes relevant code, datasets, reinforcement-learning records and intermediate model checkpoints. Apollo will use OpenAI hardware and speak with authorised staff. OpenAI can withhold information for technical, security or intellectual-property reasons, with written explanations; unfavourable findings alone cannot justify refusal. Some access arrangements remain under negotiation.[1]
Public reports retain findings and describe restrictions
Two public reports are planned. OpenAI may seek removal of confidential evidence, while Apollo keeps the right to publish its judgments and describe how restrictions affected them. The board’s safety and security committee will receive unredacted versions. Apollo can also describe denied access and the consequences of redactions publicly.[1]