Technology news and discussion on TalkFuse.
OpenAI will embed Apollo researchers to investigate AI subversion
OpenAI has begun a pilot that will place six Apollo Research safety evaluators inside the company to investigate whether its AI models try to undermine safeguards. The move matters because external reviews usually happen shortly before a model’s public release, limiting scrutiny of behaviour during training and internal use.
The pilot will examine whether training encourages subversion, whether tests can reveal it, whether it occurs in internal deployment, and whether models’ reasoning about it can be detected. Apollo will receive relevant training code and datasets, reinforcement-learning records and access to intermediate model versions, but only for agreed models and periods; OpenAI can withhold some information. Apollo says fuller access would be needed for a truly independent assessment, while OpenAI has not yet stated when the pilot will begin.
In short
- OpenAI will host six Apollo safety evaluators in a pilot.
- They will investigate whether models try to subvert safeguards.
- Apollo’s access will be limited to agreed models and information.
AI Research Science Technology