RoboRisk benchmark finds high harm rates when foundation models drive robots

Researchers have built RoboRisk, a benchmark for evaluating safety when foundation models control robots, and found the models generally show high harm rates, according to a paper on arXiv.

The benchmark's 105 tasks, 92 simulated and 13 on physical robots, span five groups: task-level refusal, plan repair, execution-time hazard, injection and sequential memory. Six models were evaluated, among them the GPT-6 family in Astra, Sol and Luna variants, plus Claude Opus 5.5, Gemini 4 Argon and Grok 4.7.

The team also proposed LEASH, the Lease-Enforced Agentic Safety Harness, which uses a decision model to review the task, each action and each irreversible operation in its visual context, with the planner and all three gates sharing one safety policy.

In simulation and on physical robots, LEASH reduced harm across all six foundation models while largely preserving task completion, adding little time and token overhead, the authors reported.