What happened
NVIDIA TensorRT-LLM contains a race condition that can be triggered by running concurrent requests, which may lead to data tampering. Fixed in v1.2.1.
Why it matters
TensorRT-LLM is a widely-deployed inference engine; a request-driven race condition that corrupts output data is a low-severity but real integrity risk in shared multi-tenant inference deployments.
Attack vector
CWE-362 race condition: an authenticated/regular user submits concurrent requests to the TensorRT-LLM inference stack, causing data tampering (low confidentiality/integrity impact, no availability impact). CVSS 3.1 6.4 (AV:N/AC:L/PR:L/UI:N/S:C).
Affected systems
TensorRT-LLM all versions prior to v1.2.1
Mitigation
Upgrade TensorRT-LLM to v1.2.1 or later (NVIDIA advisory a_id/5805).