Description

Security researchers at Wiz have discovered a critical vulnerability, dubbed NVIDIAScape (CVE-2025-23266), within Nvidia's Container Toolkit, posing a significant threat to AI cloud services utilizing shared GPU infrastructure. This flaw, publicly demonstrated at Pwn2Own Berlin, earned researchers a $30,000 reward. Nvidia has since released an advisory detailing the issue, which carries a high severity rating of CVSS 9.0. The vulnerability could enable attackers to gain elevated privileges, compromise sensitive data, manipulate systems, or disrupt essential services. The affected toolkit is crucial for running GPU-accelerated containers, particularly in cloud environments hosting demanding AI workloads. The core of the vulnerability lies in a misconfigured use of Open Container Initiative (OCI) hooks, which are intended to perform predefined actions during different stages of a container’s execution. In a shared, multi-tenant cloud setting, a malicious container could exploit this flaw to breach isolation boundaries, potentially achieving full root access to the host system. This level of access could expose confidential data and AI assets belonging to other users operating on the same hardware. Wiz researchers demonstrated that the exploit can be triggered with a straightforward, specially crafted Docker image and minimal code, underscoring the inherent fragility of container boundaries. They stressed that containers alone are not sufficient for secure isolation, especially in shared environments. To mitigate such critical threats, they recommend bolstering security with more robust techniques like virtualization to ensure comprehensive protection.