CVE-2026-41664: Samsung ONE Buffer Overflow Vulnerability

CVE-2026-41664 Overview

CVE-2026-41664 is an Integer Overflow vulnerability affecting Samsung Open Source ONE (On-device Neural Engine), a neural network compiler and runtime framework. The vulnerability exists in the memory copy size calculation logic, where processing large tensor shapes can trigger an integer overflow condition, leading to invalid memory operations.

Critical Impact

Successful exploitation could allow attackers to cause denial of service through memory corruption, with potential for limited information disclosure or integrity impacts when processing maliciously crafted neural network models.

Affected Products

• Samsung Open Source ONE versions prior to commit 1.30.0

Discovery Timeline

• 2026-04-22 - CVE CVE-2026-41664 published to NVD
• 2026-04-22 - Last updated in NVD database

Technical Details for CVE-2026-41664

Vulnerability Analysis

This vulnerability is classified under CWE-190 (Integer Overflow or Wraparound). The flaw occurs during memory copy size calculations within the Samsung ONE neural network framework when handling tensor operations with exceptionally large shapes.

When the framework processes tensor metadata, it calculates the required memory allocation size by multiplying tensor dimensions. If a crafted tensor specifies dimensions that, when multiplied together, exceed the maximum value representable by the integer type used for size calculation, the result wraps around to a smaller value. This causes subsequent memory operations to allocate insufficient buffer space, leading to out-of-bounds memory access.

The attack requires local access and user interaction (such as loading a malicious model file), which limits the attack surface. However, in scenarios where ONE processes untrusted neural network models, this vulnerability presents a significant risk.

Root Cause

The root cause is improper integer handling during tensor memory size computation. The code fails to validate whether the product of tensor dimensions exceeds the maximum integer value before allocating memory buffers. When the multiplication result overflows, it wraps to a small positive value, causing the allocation of an undersized buffer.

Attack Vector

Exploitation requires local access to the system running Samsung ONE. An attacker must craft a malicious neural network model containing tensor definitions with dimensions specifically designed to trigger the integer overflow. When a user loads or processes this model, the invalid memory operations occur.

The attack scenario typically involves:

1. Creating a neural network model with tensor shapes containing extremely large dimension values
2. Convincing a user to load the malicious model into an application using Samsung ONE
3. The framework attempts to calculate memory requirements, triggering the overflow
4. Subsequent memory copy operations operate on incorrectly sized buffers, causing crashes or memory corruption

For technical details on the fix implementation, see the Samsung GitHub Pull Request #16481.

Detection Methods for CVE-2026-41664

Indicators of Compromise

• Unexpected application crashes when loading neural network models in applications using Samsung ONE
• Memory corruption errors or segmentation faults during tensor operations
• Anomalous memory allocation patterns with unusually small buffer sizes followed by large data copy operations
• Log entries indicating tensor dimension validation failures or memory operation errors

Detection Strategies

• Monitor for crashes in processes utilizing the Samsung ONE neural network framework
• Implement file integrity monitoring for neural network model files in deployment environments
• Deploy runtime application self-protection (RASP) solutions to detect memory corruption attempts
• Audit incoming neural network models for tensor definitions with suspiciously large dimension values

Monitoring Recommendations

• Enable verbose logging for Samsung ONE framework operations during model loading
• Monitor system logs for memory-related errors in applications using neural network inference
• Implement alerting for repeated crashes in ML/AI inference services
• Track file access patterns to neural network model directories for unauthorized modifications

How to Mitigate CVE-2026-41664

Immediate Actions Required

• Update Samsung Open Source ONE to version 1.30.0 or later
• Restrict access to neural network model files to trusted sources only
• Implement input validation for neural network models before loading
• Consider sandboxing applications that process untrusted ML models

Patch Information

Samsung has addressed this vulnerability in the ONE framework. The fix is available via GitHub Pull Request #16481. Organizations using Samsung ONE should update to version 1.30.0 or apply the specific commit that resolves the integer overflow issue in memory copy size calculations.

Workarounds

• Only load neural network models from trusted and verified sources
• Implement pre-processing validation that checks tensor dimensions against reasonable bounds before passing models to the ONE framework
• Deploy applications using Samsung ONE in isolated environments with limited system privileges
• Consider implementing memory allocation limits at the operating system level for processes running ML inference