Denial of Service in NVIDIA TensorRT-LLM via gRPC Chat API

Executive Summary

CVE-2026-47470 is a vulnerability in NVIDIA TensorRT-LLM for any platform. It affects the gRPC server chat API endpoint, where an attacker can exploit it through a local attack.

The vulnerability is classified as CWE-20, which refers to improper input validation. This means the software does not properly validate input, allowing an attacker to manipulate the system in unintended ways.

A successful exploit of this vulnerability could lead to a denial of service (DoS), where the affected system becomes unavailable or unresponsive.

Useful 0 Not Useful 0

Detection Guidance

Detection of CVE-2026-47470 involves identifying the presence of the vulnerable NVIDIA TensorRT-LLM gRPC server chat API endpoint on your system or network. Since the vulnerability is related to a local attack vector, focus on systems where TensorRT-LLM is installed and running.

• Check for the installation of NVIDIA TensorRT-LLM: Use package managers or system commands to verify if TensorRT-LLM is installed. For example, on Linux, you might use 'pip list | grep tensorrt-llm' or check the installation directory.
• Identify running gRPC server processes: Use commands like 'ps aux | grep grpc' or 'netstat -tulnp | grep grpc' to check for active gRPC server processes associated with TensorRT-LLM.
• Verify the gRPC server chat API endpoint: If the gRPC server is running, you can use tools like 'grpcurl' to interact with the API and check for the vulnerable endpoint. For example, 'grpcurl -plaintext localhost:PORT list' to list available services.

Since the vulnerability is local, ensure that only trusted users have access to the system where TensorRT-LLM is running. Monitor logs for unusual activity related to the gRPC server.

Useful 0 Not Useful 0

Impact Analysis

If you are using NVIDIA TensorRT-LLM with the gRPC server chat API endpoint, this vulnerability could impact you in the following ways:

• Denial of Service (DoS): An attacker could exploit this vulnerability to crash or disrupt the service, making it unavailable for legitimate users.
• Local Attack Vector: Since the attack requires local access, an attacker with access to the system could exploit this vulnerability without needing remote access.

The CVSS base score of 6.2 indicates a medium severity, with a high impact on availability but no impact on confidentiality or integrity.

Useful 0 Not Useful 0

Compliance Impact

The impact of this vulnerability on compliance with standards and regulations depends on the context of its use:

• GDPR: If the affected system processes personal data of EU citizens, a denial of service could lead to unavailability of services, potentially violating GDPR's requirements for data availability and resilience. However, since this vulnerability does not involve data breaches or unauthorized access, the direct GDPR impact may be limited.
• HIPAA: For systems handling protected health information (PHI), a denial of service could disrupt access to critical healthcare services, potentially violating HIPAA's availability requirements. However, like GDPR, this vulnerability does not directly involve unauthorized access or disclosure of PHI.

In both cases, organizations should assess whether the vulnerability introduces risks to the availability of regulated data or services and take appropriate mitigations to maintain compliance.

Useful 0 Not Useful 0

Mitigation Strategies

To mitigate CVE-2026-47470, follow these immediate steps:

• Restrict access to the gRPC server: Limit access to the system running TensorRT-LLM to only trusted users and networks. Use firewalls or network policies to block unnecessary access to the gRPC server port.
• Disable the vulnerable gRPC server chat API endpoint: If the endpoint is not required for your use case, disable it in the TensorRT-LLM configuration to prevent exploitation.
• Apply patches or updates: Monitor NVIDIA's official channels for patches or updates addressing this vulnerability. Apply them as soon as they become available.
• Monitor for suspicious activity: Enable logging for the gRPC server and monitor logs for any unusual or unauthorized access attempts.
• Isolate the system: If possible, isolate the system running TensorRT-LLM from critical networks to minimize the impact of a potential exploit.

Useful 0 Not Useful 0