Dangerous Mind: LLMs in Robotics Create Systemic Safety Threats

The integration of AI into the physical world has collided with harsh reality. On April 21, 2026, a consortium of researchers (ETH Zurich, Stanford, UCL, etc.) released the `DESPITE` benchmark on arXiv, analyzing the safety of using LLMs for embodied planning.

The work sheds light on a critical vulnerability: even if an algorithm builds logical plans perfectly, this does not guarantee their safety when transferred to a robot chassis. By analyzing 12,279 tasks, scientists proved that foundation models are prone to generating dangerous or unpredictable sequences of actions in a real-world environment. Against the backdrop of Chinese humanoid marathons (April 19 case) and the deployment of AGIBOT platforms, the research sounds like a stop signal for the industry. Without built-in physical limiters (safety guardrails), the commercialization of Embodied AI could result in man-made disasters.

Source: ETH Zurich / arXiv