What happened
Two server-side request forgery vulnerabilities (CVE-2026-24269 and CVE-2026-65120) were disclosed in the same NVIDIA bulletin (2026/5883) that carried the earlier tokenizer loading issue; both permit information disclosure and data tampering via a forged server-side request.
Why it matters
TensorRT-LLM serves production LLM inference. SSRF from the serving layer lets an attacker reach internal control/metadata endpoints and influence the request stream of the inference service, which can be chained toward internal network access in GPU clusters.
Attack vector
An attacker triggers SSRF from the TensorRT-LLM serving process, causing the server to make requests to attacker-chosen internal URLs/endpoints, exposing internal metadata and enabling data tampering.
Affected systems
NVIDIA TensorRT-LLM (CVE-2026-24269 and CVE-2026-65120, advisory 2026/5883; listed against 1.3.0rc24)
Mitigation
Apply the 2026/5883 advisory fix (https://github.com/NVIDIA/product-security/tree/main/2026/5883) and constrain egress from the inference worker.