CVE-2026-24271 in TensorRT-LLM
Summary
by MITRE • 07/15/2026
NVIDIA TensorRT-LLM contains a vulnerability in the OpenAI-compatible inference API, where an attacker could cause allocation of GPU resources without limits or throttling. A successful exploit of this vulnerability might lead to denial of service.
VulDB is the best source for vulnerability data and more expert information about this specific topic.
Analysis
by VulDB Data Team • 07/15/2026
This vulnerability exists within NVIDIA TensorRT-LLM's OpenAI-compatible inference API implementation and represents a critical resource exhaustion flaw that can be exploited to disrupt system availability. The vulnerability stems from inadequate validation and control mechanisms surrounding GPU memory allocation requests within the API endpoint, allowing malicious actors to submit crafted requests that bypass normal resource limits and throttling controls. This weakness enables uncontrolled consumption of GPU memory and computational resources without proper bounds enforcement, creating a pathway for attackers to exhaust available GPU capacity through sustained or massive allocation requests.
The technical nature of this vulnerability aligns with CWE-400, which addresses "Uncontrolled Resource Consumption" or "Resource Exhaustion" conditions in software systems. When exploited, the flaw allows attackers to perform unauthorized memory allocation operations that circumvent the normal resource management controls built into the inference API. The OpenAI-compatible API interface provides multiple endpoints for model inference and prompt processing that can be manipulated to trigger unlimited GPU memory consumption patterns. This type of vulnerability falls under the ATT&CK technique T1499.004, specifically "Utilities: System Shutdown/Reboot," as the resource exhaustion ultimately results in denial of service conditions that prevent legitimate users from accessing the inference services.
The operational impact of this vulnerability extends beyond simple service disruption to potentially compromise entire machine availability and performance. When GPU resources are consumed without limits, it affects not only the targeted inference API but can also impact other processes running on the same system that depend on GPU compute resources. The denial of service condition manifests when legitimate users or applications cannot obtain necessary GPU memory allocations due to exhaustion, leading to failed inference requests and degraded system performance. This vulnerability is particularly concerning in production environments where TensorRT-LLM serves as part of larger AI infrastructure pipelines, as it can effectively shut down critical machine learning workloads.
Mitigation strategies for this vulnerability should focus on implementing robust input validation and resource limit enforcement within the API layer. Organizations should deploy rate limiting mechanisms that monitor and control GPU memory allocation requests per user or session to prevent unbounded consumption patterns. The implementation of proper resource quotas and monitoring systems can help detect anomalous allocation behaviors before they result in service disruption. Additionally, updating to patched versions of NVIDIA TensorRT-LLM that include proper resource management controls is essential. Network-level protections such as API gateway rate limiting, request size restrictions, and behavioral analysis tools can provide additional layers of defense against exploitation attempts. System administrators should also implement comprehensive monitoring solutions that track GPU memory utilization patterns and alert on unusual consumption spikes that may indicate exploitation attempts.