CVE-2026-24173: 
Triton Inference Server vulnerability analysis and mitigation

CVE-2026-24173 is a denial-of-service vulnerability in NVIDIA Triton Inference Server caused by an integer overflow or wraparound condition (CWE-190). An unauthenticated remote attacker can crash the server by sending a specially crafted malformed request. All versions of Triton Inference Server prior to r26.02 are affected. The vulnerability was published on April 7, 2026, with initial NVD analysis completed on April 16, 2026. It carries a CVSS v3.1 base score of 7.5 (High), assigned by NVIDIA Corporation (GitHub Advisory, NVIDIA Advisory).

The root cause is an integer overflow or wraparound (CWE-190) in NVIDIA Triton Inference Server's request handling logic. When the server processes a malformed request, an integer calculation produces an unexpected value — either wrapping around to a very small or negative number — which can trigger a server crash, consistent with a NULL pointer dereference or invalid memory access as a secondary effect. The attack requires no authentication, no user interaction, and is exploitable over the network with low complexity, making it accessible to any attacker with network reach to the server. No public proof-of-concept or detailed technical write-up has been identified at this time (GitHub Advisory, NVIDIA Advisory).

Successful exploitation results in a server crash, causing a complete loss of availability for the Triton Inference Server and any AI/ML inference services it hosts. There is no impact on confidentiality or data integrity — the vulnerability is limited to availability. In production environments where Triton serves real-time inference workloads, repeated exploitation could result in sustained service outages, disrupting dependent applications and pipelines (GitHub Advisory).

1. Reconnaissance: Identify internet-facing or network-accessible NVIDIA Triton Inference Server instances running versions prior to r26.02 using network scanning tools (e.g., Shodan, Censys, or nmap targeting default Triton ports such as 8000/HTTP, 8001/gRPC, 8002/metrics).
2. Craft malformed request: Construct a specially malformed HTTP or gRPC request designed to trigger the integer overflow condition in the server's request parsing or processing logic.
3. Send request: Transmit the malformed request to the Triton Inference Server endpoint without any authentication credentials.
4. Trigger crash: The integer overflow causes the server to crash, resulting in denial of service for all inference workloads hosted on the affected instance.
5. Repeat for sustained DoS: Resend the malformed request after any automatic restart to maintain a persistent denial-of-service condition (GitHub Advisory).

• Network: Unexpected or malformed HTTP/gRPC requests to Triton Inference Server ports (default: 8000, 8001, 8002) from unknown or untrusted source IPs; high-frequency requests with unusual payload structures.
• Logs: Triton server logs showing crash or unexpected termination events; error messages related to integer arithmetic or memory access violations around request processing.
• Process: Sudden termination or restart of the tritonserver process; core dump files generated in the Triton working directory following a crash event.
• Availability: Inference endpoints becoming unresponsive or returning connection refused errors following receipt of specific request patterns.

NVIDIA has released a fix in Triton Inference Server version r26.02; all users should upgrade to this version or later as the primary remediation (NVIDIA Advisory). As a workaround prior to patching, implement network-level access controls (firewalls, security groups) to restrict access to Triton Inference Server ports (8000, 8001, 8002) to trusted IP ranges only. Additionally, consider deploying an API gateway or reverse proxy with request validation and rate limiting in front of the Triton server to filter malformed requests before they reach the vulnerable component.

Coverage of CVE-2026-24173 has been limited to automated vulnerability tracking platforms and security news aggregators such as SecurityOnline.info and TheHackerWire, with no notable independent researcher commentary or significant social media discussion identified. NVIDIA disclosed the vulnerability through its standard customer help portal advisory process (NVIDIA Advisory).