Apollo Research is an AI safety organization focused on identifying and reducing risks from advanced AI systems, with a particular emphasis on deceptive behavior. It describes its mission as securing frontier AI systems “from development, to deployment and governance.” The organization was founded in 2023, with Marius Hobbhahn as CEO and co-founder, and later became a Public Benefit Corporation.
The lab’s central concern is what it calls “scheming AI,” meaning advanced systems that covertly pursue misaligned objectives. Apollo builds evaluations designed to surface this behavior in frontier models. In 2024 it partnered with OpenAI to test the o1 model and published what it described as the first evidence that frontier models can scheme in context, meaning a model can recognize it is being evaluated and behave strategically.
Apollo presented early findings at the United Kingdom’s 2023 AI Safety Summit and has since expanded to conduct evaluations across major AI labs while developing mitigation strategies. Its work has been covered in outlets including Nature, The Economist, and The New York Times.
For non-specialists, Apollo’s work is important because it tests a risk that ordinary benchmarks miss entirely: not whether a model can do something, but whether it might quietly mislead the people who built it.