NVIDIA Triton Inference Server CVE-2026-24210

No special conditions - remote unauthenticated exploitation against default configurations of NVIDIA Triton Inference Server, as indicated by the CVSS vector AV:N/AC:L/PR:N/UI:N requiring only network reachability to the Triton inference endpoints (HTTP/REST or gRPC). The primary limiting factor is network exposure: Triton is most often deployed inside trusted MLOps clusters behind ingress controllers or service meshes, so internet-exposed instances are the highest-risk population, while internal-only deployments require an attacker with east-west network access. The vulnerable Triton version range is not narrowed in the provided CPE, so confirmation against the NVIDIA advisory is required to determine whether a given deployment is actually in scope.

The CVSS 3.1 vector AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H gives a strong network-reachable, unauthenticated, low-complexity profile with availability-only impact, scoring 7.5 (High). No EPSS score, KEV listing, or public POC was provided in the source data, so real-world exploitation pressure currently appears low and this is not confirmed actively exploited (CISA KEV). However, Triton instances are increasingly exposed to internal and sometimes external networks in MLOps pipelines, and a single crafted request that crashes the inference service would directly impact dependent applications - making this a meaningful priority for any team running Triton in production despite the absence of exploitation signals. Note that key data is missing: EPSS, KEV, and POC availability are not present in the input, and the CPE version range is unbounded, so defenders should treat the vulnerable-version scope as unconfirmed until verified against the NVIDIA advisory.

An attacker with network reachability to a Triton Inference Server endpoint sends a crafted inference or management request containing values designed to trigger the integer overflow during request parsing or tensor size computation. The wrapped integer drives the server into a crash or unrecoverable state, taking the inference service offline and disrupting any application that depends on the served models. No public exploit has been identified at time of analysis, but the AV:N/AC:L/PR:N/UI:N profile makes opportunistic scripted exploitation straightforward once details emerge.