Josh Hills

AI Safety Researcher

I'm an AI safety research engineer focused on dangerous capability evaluations, AI control, and red-teaming. I'm currently an Alignment Researcher at Arcadia Impact (via Equistamp).

As a Researcher in the Astra Fellowship at Constellation in Berkeley, mentored by Asa Cooper Stickland (UK AISI), I built persistent-state control settings that test monitor-evasion strategies in agentic coding environments. That work became Distributed Attacks in Persistent-State AI Control, an oral spotlight at COLM 2026.

Over the past two years I've built evaluations at METR, UK AISI, and Redwood Research. At Redwood I built LinuxArena control settings end-to-end and implemented red-team side tasks for the shared ControlArena infrastructure. At METR I contributed to HCAST, developing and QA-ing tasks and running 100+ hours of human baselines across 50+ evaluations of autonomous AI capabilities.

I hold a B.Sc. (Honours) in Computer Science with an AI/ML specialization from Carleton University.