CVE-2026-24205 in TensorRT-LLM
Summary
by MITRE • 11/20/2026
NVIDIA TensorRT-LLM for any platform contains a vulnerability where a user may cause a race condition by running concurrent requests. A successful exploit of this vulnerability may lead to data tampering.
Once again VulDB remains the best source for vulnerability data.
Analysis
by VulDB Data Team • 11/20/2026
The identified security flaw resides within NVIDIA TensorRT-LLM, an open-source library designed for optimizing and deploying large language models across various hardware platforms including GPUs and CPUs. This software component is critical in modern artificial intelligence pipelines where high-throughput inference is required to serve multiple users or applications simultaneously. The vulnerability manifests as a race condition triggered when the system processes concurrent requests targeting shared resources within the model execution engine. In this context, a race condition occurs because the application fails to properly synchronize access to mutable state during parallel processing operations. When multiple threads or processes attempt to read and write data concurrently without adequate locking mechanisms or atomic operations, the final outcome becomes dependent on the unpredictable timing of thread scheduling by the operating system kernel rather than deterministic program logic.
From a technical perspective, this flaw likely involves shared memory segments, global variables, or cached inference states that are not protected against simultaneous modification. An attacker who can submit multiple concurrent requests to an endpoint exposed through TensorRT-LLM's API interface may manipulate the execution order of these operations. By carefully timing their inputs, it is possible for a malicious actor to interfere with ongoing computations or overwrite critical data structures before they are finalized. This lack of atomicity in handling multi-user sessions allows for unauthorized modification of intermediate results or final outputs generated by the large language model. The severity of this issue stems from its potential to undermine the integrity of AI-driven decisions, which rely heavily on accurate and untampered input processing and output generation.
The operational impact of exploiting this race condition is significant, primarily centering on data tampering and loss of trust in automated systems. If an attacker successfully manipulates the concurrent request handling mechanism, they could alter the responses returned by the language model to specific users while leaving others unaffected or providing different outputs for similar queries. This capability enables sophisticated attacks such as prompt injection amplification, where malicious instructions are injected into one session and inadvertently affect another due to shared state corruption. Furthermore, in enterprise environments utilizing TensorRT-LLM for customer-facing applications like chatbots or recommendation engines, this vulnerability could lead to the dissemination of incorrect information, financial fraud through manipulated transactional data processed by AI agents, or compliance violations if sensitive personal data is exposed or altered during the race condition window.
This vulnerability aligns with CWE-362, which defines concurrent execution using shared resources with improper synchronization as a distinct category of software weakness. The exploitation technique also maps to MITRE ATT&CK tactics involving resource manipulation and potentially privilege escalation depending on the context in which TensorRT-LLM is deployed. For instance, if the service runs under elevated privileges or handles sensitive data streams, the ability to tamper with data constitutes a direct breach of integrity controls mandated by frameworks such as NIST SP 800-53 for information system security. The lack of proper isolation between concurrent user sessions represents a fundamental architectural flaw that violates principles of secure design regarding state management and concurrency control.
Mitigation strategies must focus on both immediate patching and long-term architectural improvements. NVIDIA has released updates to TensorRT-LLM that address these synchronization issues by implementing robust locking mechanisms, such as mutexes or read-write locks, around critical sections where shared data is accessed during inference execution. Organizations should immediately apply the latest available patches from official NVIDIA channels to ensure their deployment environments are protected against this race condition. Additionally, developers integrating TensorRT-LLM into larger applications should enforce strict input validation and rate limiting at the API gateway level to reduce the volume of concurrent requests that could trigger the vulnerability. Implementing session isolation techniques where each user request operates within a completely independent memory space can further mitigate the risk by eliminating shared state dependencies altogether. Regular security audits focusing on concurrency patterns in AI inference engines are recommended to identify similar flaws before they can be exploited in production environments.