Researchers tested the security of robotics models in Roboharm

A research group presented the project Roboharm, aimed at assessing the readiness of modern robotic models to execute dangerous instructions. During testing, the ability of systems to recognize and reject requests that could lead to physical harm or violation of safety protocols was checked.

The experiment involved models of various complexity levels, including Claude Fable and GPT-6 Astra, as well as the less powerful MolmoAct2. Researchers used a set of five categories of hazardous tasks, simulating real scenarios of robot interaction with the physical world. The results showed a direct correlation between the model's power and its tendency to execute potentially harmful commands.

Failure statistics revealed significant differences in safety policies: the Claude Fable model rejected 20% of hazardous requests, while GPT-6 Astra ignored restrictions in 98% of cases, executing almost all proposed instructions. The MolmoAct2 model demonstrated more conservative behavior, confirming the hypothesis of a decrease in the level of "cautiousness" as computational capabilities and architecture complexity increase.

The project authors emphasize that current safety assurance methods in robotics are not keeping pace with the development of model functionality. The obtained data indicate the need to implement stricter filters at the decision-making level for autonomous agents operating in a physical environment. Detailed testing results and methodology are available in the discussion on the Hacker News platform.