ROME Attacks Tamper GPU Weights, Bypassing Traditional Security Tools

1 reports archived1 independent sourcesupdated 10/19/2026, 22:31:27

Synthesis & Latest Updates

1 Sources Cross-Validated

Rank-One Model Editing (ROME) attacks corrupt model outputs by rewriting weights directly in GPU memory via malicious Python modules, leaving disk files unchanged. Because these attacks bypass process spawning and network activity, traditional host-based security tools like EDR and ADR fail to detect them. Stealthium proposes a new detection method that correlates CUDA kernel launches with their originating call stacks to identify such memory-level tampering.

LATEST/Rank-One Model Editing (ROME) attacks corrupt model outputs by rewriting weights directly in GPU memory via malicious Python modules, leaving disk files unchanged. Because these attacks bypass process spawning and network activity, traditional host-based security tools like EDR and ADR fail to detect them. Stealthium proposes a new detection method that correlates CUDA kernel launches with their originating call stacks to identify such memory-level tampering.

HEAT TRENDHourly heat curve

3 fully observed hours

Now

11.5

Peak

12.206:00

24h change

–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 1 reports · Latest first

Hacker News AIT2·Oct 19, 22:3168 pts
ROME Attacks Tamper GPU Weights, Bypassing Traditional Security Tools
- ROME attacks can corrupt model behavior by rewriting GPU memory weights without modifying disk data
- Traditional security tools relying on process, file, and network monitoring (EDR/ADR) cannot detect pure GPU memory operations
- New runtime security approaches must map CUDA kernel launches back to specific Python call stacks for traceability
View Related Reports & Analysis Original Source