Detecting ROME attacks via CUDA kernel call stack monitoring
October 19, 2026
Rank-One Model Editing (ROME) attacks allow malicious Python modules to rewrite model weights directly in GPU memory without altering files on disk. Stealthium detects these attacks by linking CUDA kernel launches to specific Python call stacks, bypassing traditional CPU-focused EDR and ADR tools.
HOW THIS AFFECTS YOU
●
builderYou must account for GPU memory tampering that traditional host-level security tools cannot detect.
●
policyNew attack vectors require shifting security focus from file integrity to runtime kernel monitoring.