A new vulnerability dubbed “LeftoverLocals” has been uncovered that can extract sensitive data from the memory of consumer graphics processing units (GPUs).
Discovered by researchers at Trail of Bits, the flaw affects GPUs made by major vendors including Nvidia, AMD, Apple, Qualcomm, and Intel. It specifically targets GPUs used to power machine learning and large language models.
LeftoverLocals infiltrates the GPU through developer APIs and programming interfaces. It then exploits differences in how GPUs and CPUs isolate memory to steal data inside the graphics chip.
Tests showed it was possible to leak 5.5MB of data per GPU invocation from an AMD Radeon RX 7900 XT. This is enough to even completely recreate trained AI models, the researchers warned.
Two main processes are involved: a “Listener” monitors uninitialized GPU memory for canary values, while a “Writer” inserts canary values into local GPU memory. This allows attackers to exfiltrate sensitive info.
While less risky for average users, LeftoverLocals poses serious threats to cloud computing firms, researchers, and others using GPUs for AI development. Attackers could steal and replicate proprietary models or use extracted data to refine algorithms.
GPU vendors have reportedly been notified of the flaw already. But it highlights emerging risks with accessible AI tools and the need to safeguard sensitive data and models powering machine learning on commercial hardware.
Source: Trailofbits

