Hackers could poison AI models by tampering with core infrastructure (www.cybersecuritydive.com)

A recent report by security firm Eclypsium reveals that hackers could potentially compromise AI models by targeting their core infrastructure, including firmware and software that operate the computers running these models. The findings emphasize a critical vulnerability—if attackers gain access to elements like the operating system, memory controllers, or device firmware, they can bias the AI model's decisions or validate unsafe dependencies. This highlights the urgent need for enhanced security measures among AI labs and businesses utilizing such technologies. The report elaborates on how the hierarchical structure governing an AI model's runtime state presents various attack vectors. Even advanced data processing units (DPUs) employed by major chipmakers like Nvidia, AMD, Intel, and Arm can be manipulated by attackers to alter a model's operating state without changing its file integrity. Eclypsium stresses the importance of enforcing security protocols that operate independently from the AI system itself, recommending that companies maintain comprehensive inventories of their AI components and rigorously evaluate third-party software. This call for robust oversight indicates a significant recognition of the intricate security landscape surrounding AI and the necessity for proactive measures to ensure trusted operations.