PI3D Attack Enables Prompt Injection via Physical 3D Objects
August 21, 2026
The PI3D framework demonstrates how attackers can inject prompts into MLLM-driven robotics or agents by placing text-bearing physical objects in a 3D scene. The method optimizes object pose and orientation to bypass digital-only security measures and override intended tasks.
HOW THIS AFFECTS YOU
●
builderYou must account for physical-world adversarial inputs when deploying vision-language models in robotics.
●
policyThis highlights a new vulnerability surface for regulating autonomous agents in physical environments.