Businesses evaluating AI typically ask whether a model is accurate, secure and appropriate for a particular task. New research from Eclypsium, supported by NVIDIA, raises a further question: can those assurances hold if the infrastructure executing the model is compromised?
The findings accompany an Eclypsium and NVIDIA collaboration focused on addressing the emerging AI infrastructure attack surface and strengthening AI trust. Together, the companies are bringing attention to security dependencies beneath the model layer.
In a controlled demonstration, Eclypsium researchers used a compromised data processing unit, combined with permissive memory isolation, to modify a model’s weights while it ran. The original model files remained unchanged.
The potential consequences extend to the reliability of AI-generated decisions and code. An attacker able to influence a running model could manipulate outputs without leaving changes in the stored files that conventional model integrity checks examine.
The research required privileged infrastructure access and specific configuration conditions. It does not demonstrate widespread compromise. It does, however, show how an infrastructure attack could undermine confidence in an otherwise approved model.
The companies’ joint blog places infrastructure integrity within the broader effort to build trusted AI. For business leaders, the implication is that AI governance should include evidence about the security of the execution environment, alongside assessments of model performance and data handling.
As organizations give AI greater influence over business processes, the research reinforces a foundational requirement, that trust in AI outputs depends partly on the integrity of the systems producing them.
