NVIDIA TensorRT‑LLM includes a defect in its OpenAI‑compatible inference API that permits an attacker to trigger a reachable assertion failure in the sampler thread. This flaw forces the sampling process to abort, resulting in a disruption of the inference service and a denial of service. The weakness aligns with CWE‑617, indicating improper handling of recursion or control flow that can lead to assertion violations. The impact is confined to the availability of the inference endpoint, as no data exfiltration or unauthorized access is described.

The affected product is NVIDIA TensorRT‑LLM. No specific version information is included in the data; therefore, all released builds of TensorRT‑LLM that expose the OpenAI‑compatible inference API should be considered potentially vulnerable until a vendor fix is issued.

The CVSS score of 6.2 suggests a moderate severity, while the EPSS score of less than 1% indicates a low likelihood of widespread exploitation as of now. The vulnerability is not listed in the CISA KEV catalog. The likely attack vector is via the publicly exposed inference API; an attacker would need the ability to send inference requests to trigger the assertion. No prerequisite conditions are stated, so the exploitation could be performed remotely by sending crafted requests to the vulnerable endpoint.