Into the Unknown: Benchmarking AI agents on open scientific research problems