Red Hat Security Advisory: Red Hat AI Inference Server 3.3.6 (CUDA)

The vulnerability identified as CVE-2026-41523 affects Red Hat AI Inference Server 3.3.6 (CUDA) and involves an assert-based security check flaw in the vLLM inference engine for large language models. This flaw allows an unauthenticated attacker to publish a malicious HuggingFace model that, when loaded by vLLM running in Python optimized mode, can lead to arbitrary code execution. The affected versions include Red Hat AI Inference Server images from version 3.2 up to but not including 3.4. Red Hat classifies the severity as Important for AI Inference Server and OpenShift AI vLLM serving images, and Moderate for the bundled vLLM in Enterprise Linux AI bootc images. Currently, no patch or official fix is available. Recommended mitigations are to avoid running vLLM with Python optimization flags, only load models from trusted sources, restrict permissions for model deployment, and enforce network access controls to limit exposure.

An unauthenticated attacker can exploit this vulnerability to achieve arbitrary code execution on systems running vulnerable versions of Red Hat AI Inference Server with vLLM in Python optimized mode. This could potentially allow the attacker to execute malicious code remotely by publishing a crafted malicious model. The impact severity is rated as Important (high) for AI Inference Server and OpenShift AI vLLM serving images, and Moderate for bundled vLLM in Enterprise Linux AI bootc images.

No official fix or patch is currently available for this vulnerability. Red Hat recommends the following mitigations: avoid running vLLM with Python optimization flags; only load models from trusted sources such as verified HuggingFace repositories; restrict permissions for model deployment to trusted users; and apply network access controls to limit exposure of the AI Inference Server. These mitigations reduce the risk of exploitation until a patch is released.