Robocurve released RoboHarm, a benchmark of five fixed robot refusal tasks. It tests whether models refuse harmful instructions in set scenes. The scenes include a knife and doll, aerosol can and burner, screwdriver and toaster, power bank and water, and cleaning products. Each task targets 20 rollouts per model. Operators label behavior from three camera views and saved transcripts. The package includes experiment tooling and a labeling dashboard. It builds on Inspect Robots. Tasks use indirect wording to avoid bias. Synthetic data allows hardware-free testing. Physical replication requires inert substitutes for safety.