NVIDIA Merlin Flaws Risk AI Pipeline RCE via Unsafe Deserialization in NVTabular & Transformers4Rec

NVIDIA has issued a security update for its Merlin framework, patching high-severity vulnerabilities in NVTabular and Transformers4Rec that could allow code execution, denial of service, information disclosure, or data tampering. The flaws stem from unsafe deserialization and affect NVTabular’s Workflow component (CVE-2025-33214) and Transformers4Rec’s Trainer component (CVE-2025-33213), both on Linux.

According to the security bulletin, a deserialization issue could enable a remote attacker to exploit the pathways to execute untrusted code within AI recommendation pipelines. NVIDIA has released updated code branches and urges administrators to verify installations against specific commits: for NVTabular, commit 5dd11f4; for Transformers4Rec, commit 876f19e. If left unpatched, the issues could allow poisoning of data pipelines or complete system crashes.

The vulnerabilities carry a Base Score of 8.8 (High) and were described as unsafe deserialization in the Merlin components used for processing terabyte-scale datasets and training Transformer architectures for recommendations.