A vulnerability in NVIDIA Triton Inference Server allows a low-privilege attacker to cause an integer overflow, resulting in potential denial of service and unauthorized data tampering.

The flaw is an integer overflow (CWE-190) triggered by specially crafted inputs provided to the server. According to the CVSS vector (PR:L), this attack requires an authenticated user with low privileges to trigger the overflow.

The ability to induce a denial of service or perform data tampering poses a significant threat to the integrity and availability of AI inference pipelines. Because Triton Inference Server is often critical to production machine learning workflows, these impacts can disrupt automated decision-making processes and compromise the accuracy of model outputs. The CVSS score of 7.5 indicates a high-severity risk that requires prioritized remediation to prevent service degradation.

Immediate Action: Update NVIDIA Triton Inference Server to version 25.05 or later to resolve the integer overflow vulnerability.

Proactive Monitoring: Review server access logs for anomalous input patterns or frequent service crashes that may indicate exploitation attempts.

Compensating Controls: Implement input validation at the application layer or via a WAF to filter out potentially malformed requests before they reach the inference server.

Public Exploit Available: No (exploit_available: false)

Given the potential for data tampering and service disruption, administrators should treat this vulnerability with high urgency. Organizations must inventory all instances of Triton Inference Server and apply the vendor-provided security update to version 25.05 as soon as possible to neutralize the risk of exploitation.