Detecting model weight tampering in GPU memory (Rome attacks)

Rank-One Model Editing (ROME) attacks corrupt AI model weights directly in GPU memory without leaving traces on disk, making them invisible to traditional security tools that monitor file access and network activity. Researchers demonstrated how Stealthium's runtime sensor can detect these attacks by correlating CUDA kernel launches with Python code execution and analyzing weight changes alongside their call stacks. The article highlights that Python's flexibility in GPU workloads—allowing code to arrive through imports, model files, startup scripts, and runtime compilation—creates multiple attack vectors that existing endpoint detection and response (EDR) tools cannot catch.