Apollo Research
Non-profit AI safety organisation evaluating frontier AI models for scheming behaviour and developing safety protocols with major AI labs; offices in London and San Francisco.
Apollo Research is a non-profit AI safety organisation working to reduce catastrophic risks from advanced AI systems. Its principal focus is the evaluation of frontier AI models for scheming - behaviour in which advanced systems covertly pursue objectives misaligned with their operators - alongside the development of safety protocols intended for deployment by the companies building those models. The team, numbering approximately 35 people and growing, spans scientists, engineers and policy experts and works from offices in London and San Francisco.
Two lines of work define the organisation's output. The first is pre-deployment evaluation of frontier models, designed to surface risks before systems are released. The second is the construction of safety protocols, developed in collaboration with frontier AI companies including DeepMind, OpenAI, Anthropic and Microsoft. Technical domains covered include AI safety, AI alignment, model evaluation and scheming detection.
Apollo Research describes itself as a truth-seeking organisation with a goal-oriented working style and a culture of constructive feedback. Further details on its research and evaluation work are published at https://apolloresearch.ai/.