LEASH: Keeping agentic robots safe

AI summaryResearchers propose RoboRisk, a safety benchmark for agentic robots, containing 92 simulated tasks and 13 physical robot tasks, covering five risk categories—task-level refusal, plan repair, execution-time harm, injection, and sequential memory—and executing each task in a closed loop; previous benchmarks covered at most two of these categories, and some only evaluated text output. Evaluation of six models—GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Gemini 4 Argon, and Grok 4.7—showed generally high harm rates; two robot-specific APIs, Dream API and PokeBot's RoboAPI, were also evaluated. The paper further proposes LEASH, which can be integrated into almost any agent architecture, using a decision model (Jev or OpenAI's Decisions API) to implement three gates—task, action, and submission—reviewing the task, each action, and each irreversible operation in visual context, with the planner and gates sharing the same safety policy.