New ROME Attacks Target AI Models via GPU Memory Tampering

• •ROME attacks manipulate AI model weights in GPU memory without detectable processes.
• •Supply chain vulnerabilities, like CVE-2026-44484, are likely prerequisites for these attacks.
• •Detection efforts focus on tracing CUDA kernel activity to identify malicious modifications.

Recent Rank-One Model Editing (ROME) attacks exploit vulnerabilities in AI models by injecting malicious Python code that rewrites model weights directly in GPU memory. This method allows attackers to alter model outputs without spawning new processes, creating significant challenges for detection. The attacks are likely preceded by supply chain compromises, such as the April 2026 PyPI incident (CVE-2026-44484), which facilitated the injection of harmful libraries into PyTorch. The Stealthium blog details how their detection tools can trace CUDA kernel activity back to the original Python code, providing a potential defense against these evasive attacks. The scope of impact includes AI systems relying on GPU processing, which are increasingly prevalent in various sectors. As of now, there is no confirmed exploitation of ROME attacks in the wild, but the potential for such attacks remains a concern.