CVE-2025-55551: PyTorch torch.linalg.lu DoS Vulnerability

CVE-2025-55551 Overview

A Denial of Service (DoS) vulnerability has been identified in PyTorch v2.8.0, specifically within the torch.linalg.lu component. The vulnerability is triggered when performing slice operations on the output of the LU decomposition function, allowing attackers to cause application crashes and service disruption. This issue affects machine learning applications and services that rely on PyTorch's linear algebra functionality.

Critical Impact

Attackers can remotely trigger a Denial of Service condition in PyTorch applications by exploiting improper handling of slice operations on torch.linalg.lu outputs, causing service unavailability without requiring authentication.

Affected Products

• Linux Foundation PyTorch v2.8.0
• PyTorch Python package versions prior to patched releases
• Applications and services built on vulnerable PyTorch versions

Discovery Timeline

• 2025-09-25 - CVE-2025-55551 published to NVD
• 2025-10-03 - Last updated in NVD database

Technical Details for CVE-2025-55551

Vulnerability Analysis

This vulnerability resides in PyTorch's linear algebra module, specifically within the torch.linalg.lu function which performs LU decomposition on matrices. LU decomposition is a fundamental operation in numerical computing used to solve linear equations, calculate determinants, and invert matrices. The flaw manifests when subsequent slice operations are performed on the decomposition results, leading to resource exhaustion or unhandled exceptions that crash the application.

The vulnerability can be exploited remotely over a network without requiring any authentication or user interaction. While the confidentiality and integrity of the system remain unaffected, the availability impact is significant as successful exploitation results in complete denial of service for the affected application.

Root Cause

The root cause is categorized under CWE-400 (Uncontrolled Resource Consumption). The torch.linalg.lu component fails to properly validate or handle certain slice operations on its output tensors. This improper resource management allows attackers to craft specific inputs that trigger excessive resource consumption or cause the application to enter an unrecoverable error state, resulting in service disruption.

Attack Vector

The attack can be executed remotely over the network. An attacker can exploit this vulnerability by sending specially crafted requests to applications that process user-supplied data through PyTorch's torch.linalg.lu function. The attack requires no privileges or user interaction, making it particularly accessible to malicious actors. When a vulnerable application processes the malicious input, the slice operation on the LU decomposition output triggers the DoS condition.

The vulnerability mechanism involves improper handling of tensor slice operations following LU decomposition. When specific slice parameters are applied to the output matrices, PyTorch fails to properly manage the operation, leading to application crashes. For detailed technical analysis and proof-of-concept information, refer to the GitHub Gist Resource and the GitHub Issue Discussion.

Detection Methods for CVE-2025-55551

Indicators of Compromise

• Unexpected application crashes or service restarts in PyTorch-based applications
• Error logs showing exceptions related to torch.linalg.lu or tensor slice operations
• Increased memory consumption or CPU usage patterns preceding application failures
• Multiple failed requests or abnormal input patterns targeting linear algebra endpoints

Detection Strategies

• Monitor application logs for stack traces involving torch.linalg.lu and slice operations
• Implement anomaly detection for unusual input patterns to machine learning inference endpoints
• Track application availability metrics to detect recurring crash patterns
• Use runtime application self-protection (RASP) to detect exploitation attempts

Monitoring Recommendations

• Enable verbose logging for PyTorch applications processing external inputs
• Set up alerting for application crashes with core dumps referencing the linalg module
• Monitor system resource utilization for abnormal spikes before service failures
• Implement health check endpoints to quickly detect service unavailability

How to Mitigate CVE-2025-55551

Immediate Actions Required

• Audit all applications using PyTorch v2.8.0 to identify exposure to this vulnerability
• Implement input validation and sanitization before passing data to torch.linalg.lu
• Consider isolating PyTorch processing in sandboxed environments to limit DoS impact
• Monitor the official PyTorch security channels for patch releases

Patch Information

As of the last update on 2025-10-03, no official vendor patch has been announced. Organizations should monitor the GitHub Issue Discussion for updates on fix availability. Once a patch is released, prioritize upgrading PyTorch installations, especially in production environments handling external input.

Workarounds

• Implement rate limiting on endpoints that process data through torch.linalg.lu
• Add try-catch exception handling around LU decomposition and slice operations to prevent complete application crashes
• Deploy application behind a web application firewall (WAF) to filter potentially malicious requests
• Consider using alternative linear algebra implementations for critical applications until a patch is available

# Example: Add resource limits for PyTorch applications
# Limit memory to prevent runaway resource consumption
ulimit -v 4194304  # 4GB virtual memory limit

# Monitor PyTorch process for abnormal behavior
watch -n 5 'ps aux | grep python | grep pytorch'