Apollo Research, an organization established in 2023, proposes a shift in how AI models are tested. The group advocates for 'embedded evaluations,' where independent testers gain continuous access to models during the training phase rather than performing checks only before a public release.
The organization detailed this approach in a series of blog posts released between September 30 and October 5, 2026. Apollo argues that current testing methods miss critical development stages, failing to identify risks like scheming or reward-seeking behavior that emerge during training.
CEO Marius Hobbhahn presented this proposal to the US Senate Homeland Security Subcommittee in early October 2026. He argued for mandatory independent evaluations, suggesting that voluntary cooperation from AI labs is insufficient to address the risks associated with frontier model development.
Apollo has previously collaborated with major AI labs, including research with OpenAI on reward-seeking behavior. The new proposal aims to address the challenge of limited access to training processes, which are currently among the most protected operations in the technology sector.
