A new GPU vulnerability, “LeftoverLocals,” has been discovered by security researchers, posing a threat to the security of data stored in GPU memory. According to foreign media wccftech, major GPU vendors such as Nvidia, Apple, AMD, and Qualcomm are currently affected by this vulnerability.
The “LeftoverLocals” vulnerability has the ability to extract critical data from GPU memory, impacting both consumer applications and GPUs used in LLM and ML models. Security researchers at Carnegie Mellon University are closely tracking this vulnerability, and relevant information has been shared with affected major GPU vendors.
Running a 7 billion parameter model on an AMD Radeon RX 7900 XT, the LeftoverLocals vulnerability leaked approximately 5.5 MB of data per GPU call. This speed of data leakage is significant enough to recreate a complete model, posing a high risk in the field of AI, especially for companies that focus on training LLM. LeftoverLocals relies on how the GPU isolates memory, which is different from the CPU framework, allowing an attacker who gains access to the GPU share through the programmable interface to steal memory data.
While ordinary consumers may not have much to worry about, industries involved in cloud computing or inference may find the security of LLM and ML frameworks questionable. This vulnerability highlights the importance of regular security updates and the need for vigilance in the ever-evolving landscape of technology.