NVIDIA DALI: out-of-bounds write in pipeline input handling corrupts memory

Impact

An uncontrolled out-of-bounds write in DALI corrupts memory in the process that loads training or inference data. NVIDIA lists denial of service, information disclosure and data tampering for this id - notably not code execution, unlike the stack overflow in the same bulletin. For an operator the realistic outcome is a crashed or silently corrupted data pipeline: a training run that produces a bad model is harder to notice than one that dies. It is one of four separate ids in bulletin 2026/5900, each a different memory-safety mechanism.

Who can reach it

Local, authenticated low-privileged user, with user interaction (CVSS AV:L/AC:L/PR:L/UI:R) - requires getting a crafted input into a pipeline someone runs.

What to do

Upgrade DALI to the fixed version in NVIDIA bulletin 2026/5900, rebuild the images that pin it, and restart affected jobs. One package bump covers all four ids in the bulletin, so do them in a single pass. The NVD record does not state the fixed version.

References

Related entries

• NVIDIA DALI: out-of-bounds reads while decoding pipeline input lead to disclosure and crashesCVE-2026-65147 · NVIDIA DALI (data pipeline file/media decoding)High
• NVIDIA DALI: stack-based buffer overflow in pipeline input handling allows code executionCVE-2026-65148 · NVIDIA DALI (data pipeline file/media decoding)High
• JupyterHub: A user granted limited access can escalateCVE-2024-41942 · JupyterHubHigh
• Weaviate: Crafted entry name with an absolute pathCVE-2025-67818 · WeaviateHigh
• llama.cpp (`gguf_init_from_file`): Use of uninitialized heap variableCVE-2024-32878 · llama.cpp (`gguf_init_from_file`)High
• MLflow: basic-auth app leaves tracing and assessment endpoints without permission checksCVE-2025-15381 · MLflow Tracking Server (basic-auth app: tracing and assessment endpoints)High