Luthien attended ControlConf, the inaugural conference dedicated to AI control, held March 27–28, 2025 in Marylebone, London.
About ControlConf
ControlConf is a collaboration between Redwood Research, FAR.AI, and the UK AI Security Institute. It brought together researchers, engineers, and organizations working on AI control—the study of techniques that mitigate security risks from AI even if the AI itself is trying to subvert them.
The conference drew participants from across the AI safety community:
- Researchers actively working on AI control at frontier labs, government departments, nonprofits, and academia
- AI researchers interested in learning more about AI control but lacking direct experience
- Interdisciplinary experts approaching AI control from information security, policy, and related backgrounds
Conference Objectives
The conference focused on three main goals:
- Coordination: As many people began exploring AI control, the conference fostered discussion on core strategic questions like “How ambitious should control strategies be?” and “What research is most pressing right now?”
- Knowledge sharing: Helping participants—many new to the field—understand the state of AI control research and practice.
- Interdisciplinary discussion: Exploring AI control issues with experts from AI policy, information security, and related areas to bring broader perspectives into the work.
Why We Went
Luthien's founder Jai Dhyani attended ControlConf to get feedback on Luthien's approach, connect with more AI control experts, and identify where Luthien can have the most impact on the neglected/important/tractable frontier.
The conference validated a core part of our thesis: AI control is not just a theoretical concern for future scheming AIs—it's a practical need for today's deployed systems. The conversations at ControlConf directly informed our 21 Points framework, published two weeks later.
Originally published on luthienresearch.org.
← Back to blog