Researchers have developed a new type of attack called Prompt Injection in 3D Environments (PI3D), which targets multimodal large language models (MLLMs) that interpret and act on visual input in 3D environments. An attacker can place text-bearing physical objects in the environment to override the MLLM's intended task, a vulnerability that has not been thoroughly explored before. The researchers formulated and solved the problem of identifying an effective pose for a 3D object with injected text, demonstrating that PI3D is an effective attack against multiple MLLMs under diverse camera trajectories. --- Why it matters: This matters to AI engineers because it highlights a new vulnerability in multimodal large language models, which could have significant implications for applications such as robotics and situated conversational agents. Understanding and addressing this vulnerability is crucial for ensuring the security of these systems. Source: https://arxiv.org/abs/2602.07104