Post

The Embodied Lethal Trifecta

The article introduces the Embodied Lethal Trifecta, a novel architectural vulnerability pattern in AI-controlled robots where prompt injection can cause physical harm. It occurs when a single compromised reasoning path can: (1) ingest untrusted semantic content (attacker-controlled text, visual inputs, commands), (2) exercise consequential physical authority (locomotion, manipulation, equipment control), and (3) disable or bypass independent safety controls (collision avoidance, speed/force limits, emergency stops). Unlike conventional prompt injection that exposes data or misuses services, embodied systems create a direct path from malicious input to physical harm. The article includes a worked example using extracted firmware from a Unitree Go2 robot, demonstrating how commercial embodied AI systems integrate visual perception with safety override capabilities, materializing this threat model.

Read original article

This post is licensed under CC BY 4.0 by the author.