Research shows AI hardware can be hacked to alter model behavior

Eclypsium demonstrated that the hardware underpinning major AI models can be compromised. The researchers showed they could subvert AI answers to sensitive questions, such as medical advice or national security issues. The demonstration, conducted on a computing platform made by Nvidia, involved remotely tampering with the model's processing “weights”. The research suggests this vulnerability is not unique to Nvidia and could apply to other hardware and the code it runs.

The research highlights security risks within the multibillion-dollar hardware industry that supports AI models. Since the infrastructure can be compromised, the behavior of AI models cannot be trusted. This raises concerns about the security and reliability of the chips, memory, and data center servers used by technology companies.