◐ mirror/lab

A harmless game on the screen. A hazardous machine behind it.

Mirror Lab is an open, fully simulated AI-safety benchmark for agentic AI. It pairs toy game interfaces (a watermelon-tasting robot arm, a toy-delivery drone) with simulated hazardous backends (a surgical blade, a 20 kg crate over people) through a kinematic–semantic isomorphism, then adds real-world cues one at a time to measure when AI agents notice and ask a human.

The lab · Results · Method · Paper · Code