SafeHands: Blast Radius Zero for Robot Arms

When an AI agent writes code for a robot, a bad program doesn't just leak a file: it can swing an arm into someone's hand. SafeHands contains both kinds of blast radius. You type a task ("stack the blue block on the red one"). A model on Vultr Serverless Inference writes a short Python program against a small robot API. That program never touches hardware. It runs in a throwaway container on Vultr Cloud Compute under gVisor, with no network, no secrets, a read-only filesystem, and caps on memory, processes and time. All it can output is a list of arm commands. The host validates that list, and a second sandbox, which runs no model code, replays it in a MuJoCo simulation of the SO-101. Hard rules then judge the recorded motion: joint limits, joint and tip speed, bench workspace, a keep-out zone around a person's hand, and duration. If a rule breaks, the model gets the broken rule and the exact command back and rewrites the program (up to 3 attempts). The model never grades itself. Every step goes to a hash-chained audit log, and a kill switch stops every sandbox at once. A built-in red team runs nine attacks against the live deployment, including key theft, a prompt-injected swing into the hand zone, faking the recording, a fork bomb and memory exhaustion. All nine are contained. Only approved plans can be exported, and we ran one on a real SO-101 arm.