An integer overflow vulnerability in NVIDIA Triton Inference Server allows unauthenticated remote attackers to trigger a denial-of-service, posing a significant risk to service availability.

This vulnerability involves an integer overflow (CWE-190) within the server, which can be triggered by an unauthenticated attacker via a network-based attack vector.

Successful exploitation results in a denial-of-service, rendering the inference server unresponsive. Given the CVSS score of 7.5 (High), this vulnerability poses a substantial risk to production environments relying on real-time model inference, potentially leading to critical operational downtime.

Immediate Action: Upgrade all instances of NVIDIA Triton Inference Server to version r26.03 or later.

Proactive Monitoring: Monitor server logs for unexpected crashes, restarts, or abnormal memory usage patterns indicative of overflow attempts.

Compensating Controls: Implement network-level rate limiting and ensure the server is isolated from public-facing networks to restrict unauthorized access to the inference endpoint.

Public Exploit Available: Unknown

The severity of this vulnerability, combined with its network-accessible nature, necessitates immediate action. Administrators should prioritize patching to version r26.03 across all affected infrastructure to eliminate the risk of service disruption.