Edge AI systems - from smart cameras and drones to autonomous machines - rely on neural processing units to run the object detection models that determine what the system perceives. Voltage fault injection can manipulate that perception without touching the model, firmware or software stack: objects disappear, phantom detections appear and the system continues operating, completely unaware anything is wrong. Exploring the fault parameter space manually is slow, so the question becomes whether an LLM agent can take over the search autonomously.

In this session, led by Jasper van Woudenberg and Rajesh Velegalati of Keysight, you will learn:

• How precisely timed voltage glitches on a commercial NPU cause object hallucinations, false detections and confidence score manipulation;
• How LLM-driven agents autonomously analyze detection outputs, adapt glitch parameters and optimize fault conditions overnight;
• Why neural accelerators represent an emerging and largely unaddressed hardware attack surface in edge AI deployments.