NVIDIA AI Stack Hit by Significant Security Flaw Due to Triton Vulnerabilities

NVIDIA’s Triton Inference Server has critical vulnerabilities that risk full AI stack compromise. Discovered by cybersecurity firm Wiz, these flaws in the Python backend can allow remote attackers to control AI infrastructure. The main issue arises from a malformed request that reveals internal shared memory identifiers, enabling unauthorized access to server memory. This improper validation allows attackers to exploit IPC vulnerabilities and execute remote commands, leading to potential data corruption and manipulation of AI responses.

To mitigate these risks, NVIDIA released an urgent patch (version 25.07), addressing vulnerabilities identified as CVE-2025-23319, CVE-2025-23320, and CVE-2025-23334. Organizations using Triton, including major enterprises like Amazon and Microsoft, must upgrade to ensure their AI deployments remain secure. While no current exploitations are reported, the potential for attacks persists. Wiz offers tools like Wiz Sensor and Wiz Code for vulnerability detection and remediation to safeguard AI environments effectively.

Source link