Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-65147 Overview
CVE-2026-65147 affects NVIDIA Data Loading Library (DALI), a GPU-accelerated library for data loading and preprocessing in deep learning workflows. The vulnerability allows an attacker to trigger an uncontrolled out-of-bounds read [CWE-125] within the library. Successful exploitation can lead to information disclosure and denial of service.
The flaw requires local access, low privileges, and user interaction, which limits remote exploitation. However, exploitation impacts confidentiality, integrity, and availability within affected systems running DALI workloads.
Critical Impact
An attacker with local access can read memory outside allocated buffers in NVIDIA DALI, exposing sensitive data and potentially crashing deep learning data pipelines.
Affected Products
- NVIDIA DALI (Data Loading Library)
- Deep learning training environments using DALI for preprocessing
- GPU-accelerated inference pipelines leveraging DALI
Discovery Timeline
- 2026-10-13 - CVE-2026-65147 published to NVD
- 2026-10-15 - Last updated in NVD database
Technical Details for CVE-2026-65147
Vulnerability Analysis
The vulnerability is classified under CWE-125 Out-of-Bounds Read. NVIDIA DALI processes input data such as images, video, and tensors through optimized GPU operators. When specific input conditions are met, the library reads memory outside the bounds of an allocated buffer.
Out-of-bounds reads in native code libraries like DALI can disclose adjacent process memory. In machine learning pipelines, this memory may contain model weights, training data, or other sensitive artifacts. The read access can also cause segmentation faults, terminating the host process and interrupting training or inference jobs.
The CVSS scope is unchanged, meaning the impact remains within the vulnerable component. All three impact metrics are rated High, indicating meaningful consequences for confidentiality, integrity, and availability of DALI-dependent workloads.
Root Cause
The root cause involves missing or insufficient bounds checking when DALI parses or processes input data. The library accesses memory beyond the allocated buffer size, violating memory safety guarantees. The specific operator or parsing routine responsible has not been publicly detailed by NVIDIA.
Attack Vector
Exploitation requires local access with low privileges and user interaction. An attacker supplies crafted input data processed by a DALI pipeline. This may occur when a user loads a malicious dataset, model artifact, or media file into a workflow that invokes DALI operators. The attacker does not need administrative rights but depends on a user triggering the processing operation.
No public proof-of-concept exploit is available. The EPSS score indicates low near-term exploitation probability, consistent with the local attack vector and required user interaction.
Detection Methods for CVE-2026-65147
Indicators of Compromise
- Unexpected crashes or segmentation faults in processes importing the DALI Python module or linking its native libraries.
- Deep learning training or inference jobs terminating abnormally after loading new datasets or input files.
- Memory access violations logged by the operating system kernel involving libdali or related NVIDIA runtime binaries.
Detection Strategies
- Inventory systems with NVIDIA DALI installed by scanning Python environments for the nvidia-dali package and tracking installed versions.
- Compare installed versions against the fixed release referenced in NVIDIA Security Update 5900.
- Monitor process behavior for abnormal memory read patterns or crashes in GPU workloads originating from untrusted input data.
Monitoring Recommendations
- Enable core dump collection on machine learning hosts to capture crash evidence from DALI processes for triage.
- Log the provenance of datasets and input files loaded into DALI pipelines, especially those from external or shared sources.
- Alert on repeated process termination events in training clusters that correlate with specific input data batches.
How to Mitigate CVE-2026-65147
Immediate Actions Required
- Identify all hosts, containers, and workflows that import or depend on NVIDIA DALI.
- Apply the patched DALI version referenced in NVIDIA Security Update 5900 as soon as it is available in your environment.
- Restrict access to DALI-dependent systems to trusted users and reject input data from untrusted sources until patched.
Patch Information
NVIDIA has published a security bulletin tracked as advisory 5900. Refer to the NVIDIA product security repository for the specific fixed version and upgrade instructions. See also the NVD CVE-2026-65147 Detail and the CVE-2026-65147 Record.
Workarounds
- Validate and sanitize all input data processed by DALI pipelines, rejecting malformed or unexpected formats before they reach the library.
- Isolate DALI workloads in sandboxed containers with least-privilege service accounts to limit the blast radius of information disclosure.
- Disable or avoid DALI operators that process untrusted inputs until the patch is applied.
# Check installed DALI version in your Python environment
pip show nvidia-dali-cuda120 | grep -i version
# Upgrade to the patched release once published by NVIDIA
pip install --upgrade nvidia-dali-cuda120
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

