London ControlConf 2025

A collaboration between Redwood Research, FAR.AI, and the UK AI Security Institute, ControlConf brought together individuals interested in AI control (introduced here), including:

• Researchers actively working on AI control at frontier labs, government departments, nonprofits, academia, and elsewhere
• AI researchers who want to learn more about AI control but lack direct experience
• People approaching AI control from non-AI backgrounds—such as information security professionals, policy researchers, and others who are interested in strategies to mitigate catastrophic misalignment risk

Our main objectives were:

1. Coordination
   Many people are now exploring AI control, so we aimed to foster discussion on core strategic questions like “How ambitious should control strategies be?” and “What research is most pressing right now?”
   
2. Knowledge Sharing
   As many individuals are new to the field, or have yet to engage with it directly, this conference aimed to help participants understand the state of the field.
   
3. Interdisciplinary Discussions
   We explored AI control issues with experts from related areas, including AI policy and information security, to bring broader perspectives into the discussion.

Conference content featured a mix of presentations, demos, panel discussions, opportunities for one-on-one conversations, and structured breakout sessions.