CVE-2026-16497

NVIDIA Triton Inference Server for Linux contains an excessive iteration flaw (CWE-834), in which a code loop can iterate far more times than intended when processing attacker-controlled input. Because the vulnerability is reachable over the network with no privileges or user interaction required (per the CVSS vector), a remote, unauthenticated attacker can trigger it by sending crafted requests to an exposed Triton inference endpoint. A successful exploit causes high impact to availability only, allowing the attacker to deny service to the inference server; confidentiality and integrity are not affected. Anyone running NVIDIA Triton Inference Server on Linux is potentially affected, most notably operators exposing Triton's HTTP/gRPC endpoints to untrusted networks. No public proof-of-concept, KEV listing, or confirmed in-the-wild exploitation is known, and EPSS suggests a low (0.4%) 30-day exploitation probability.

What to do: Upgrade Triton Inference Server to the patched release identified in NVIDIA's security bulletin for CVE-2026-16497, as no fixed version is specified in the available data. Until patching, restrict network access to Triton's HTTP/gRPC inference endpoints to trusted clients via firewall rules or a reverse proxy, since the flaw is exploitable by unauthenticated network requests. Note that this is a denial-of-service issue, so prioritize patching for inference services with strict availability requirements.