Jun 2026
AI Control Hackathon
A week-long research hackathon on AI control in Montréal. I was one of the co-organizers, and I built the starting point everyone worked from.
- Result
- 92 signups, 20 teams formed, 8 projects submitted
- Built with
- Control Tower, Inspect AI
AI control is the research field of getting useful work out of an AI system you don't trust, while limiting the harm it could do if it turned against you. The AI Control Hackathon ran from June 11 to 18, 2026, with sponsorship from Redwood Research (who invented the field), NVIDIA, Rootly and Rubric Labs. I co-organized it with Horizon Omega and HackOS, and we put it together in under a week.
I also introduced the other two organizers to each other; they hadn't met before. My main job, though, was the preparation. I wrote most of the starter repository, so that nobody began from a blank slate, but had a starting point for each of the five tracks (incl. commands for a first honest and attack run, an example monitor and protocol to improve on, a policy template for the red team, and a template for new side tasks). During the week I answered participants' questions, fixed bugs as they turned up, and coordinated with our contact at Redwood Research.
I also made the slides that played in the background during the opening ceremony.
92 people signed up on Luma, 20 teams filled in the team formation form, and 8 submitted a project by the end of the week. It was the first hackathon of its kind to be organized independently, and afterwards some MIT students got in touch to ask how to run a research hackathon like it for MIT AI Alignment.
The organizers' retrospective covers the winning teams and what they built.