A poisoned Python library rewrote model weights in GPU memory, and nothing on disk changed
Detecting model weight tampering in GPU memory (Rome attacks)
Rank-One Model Editing (ROME) attacks patch a model's weights directly in GPU memory, so the checkpoint on disk stays byte-for-byte identical and EDR tools watching files, processes, and sockets see an ordinary inference job. Stealthium reproduced the attack in its lab and caught it by tying each CUDA kernel launch back to the Python caller stack, flagging both the suspicious caller and kernels a clean run never launches.
Models can be corrupted about a specific topic without a single byte on disk changing. There is nothing to scan, no process to catch, no signature to hunt and match.