The vulnerability is a server-side request forgery in NVIDIA TensorRT-LLM’s multimodal media fetching functions. This flaw allows an attacker who can reach the service over the network to compel the server to make HTTP requests to arbitrary destinations. If successfully exploited, the server could become overloaded, resulting in denial of service, or it could retrieve and expose sensitive information.

NVIDIA TensorRT-LLM running on Linux is affected, as noted by the vendor’s identifiers. No specific versions are listed, so all installations of TensorRT-LLM remain potentially vulnerable until a patch is applied or a more restrictive configuration is enforced.

The CVSS score of 6.8 indicates a moderate to high severity, confirming that the flaw allows remote exploitation. The EPSS score below 1% suggests that, while the flaw exists, the overall likelihood of real-world exploitation is low; however, the fact that it can lead to denial of service and information disclosure is a concern. The flaw is not currently listed in CISA’s KEV catalog, but it represents a notable vector for attackers seeking SSRF payloads. The attack would require the attacker to send a crafted request to the TensorRT-LLM service from outside the trusted network, a path that is easily available in many production deployments.