CVE-2024-31580: PyTorch Heap Buffer Overflow DoS Flaw

CVE-2024-31580 Overview

A heap buffer overflow vulnerability was discovered in PyTorch before version v2.2.0 within the /runtime/vararg_functions.cpp component. This vulnerability allows attackers to cause a Denial of Service (DoS) via a crafted input that triggers an out-of-bounds read operation in the heap memory.

Critical Impact
Attackers can exploit this heap buffer overflow to crash PyTorch applications, potentially disrupting machine learning pipelines, AI inference services, and research environments that depend on PyTorch.

Affected Products

Linux Foundation PyTorch versions prior to v2.2.0
PyTorch Python packages (cpe:2.3:a:linuxfoundation:pytorch:*:*:*:*:*:python:*:*)
Applications and services utilizing vulnerable PyTorch runtime components

Discovery Timeline

2024-04-17 - CVE-2024-31580 published to NVD
2025-06-10 - Last updated in NVD database

Technical Details for CVE-2024-31580

Vulnerability Analysis

This vulnerability is classified as CWE-122 (Heap-based Buffer Overflow). The flaw exists in PyTorch's JIT (Just-In-Time) runtime, specifically within the tupleConstruct function in vararg_functions.cpp. The vulnerability was identified through fuzzing efforts targeting the torch JIT lite interpreter, which revealed a read-heap-buffer-overflow condition.

The root cause stems from insufficient bounds checking when constructing tuples from the stack. When the tupleConstruct function is called with a num_inputs parameter that exceeds the actual stack size, the function attempts to read beyond the allocated heap buffer, resulting in an out-of-bounds memory access.

Root Cause

The tupleConstruct function in /runtime/vararg_functions.cpp did not validate that the requested number of inputs (num_inputs) was within the bounds of the available stack size before attempting to read from the stack. This missing validation allowed attackers to supply a crafted input that specified more inputs than were actually present, causing the function to read past the end of the heap-allocated stack buffer.

Attack Vector

The attack requires local access to execute. An attacker can craft malicious input that specifies an invalid number of inputs to the tuple construction operation. When the PyTorch JIT runtime processes this input, it attempts to read from memory locations outside the allocated buffer boundaries, triggering the heap buffer overflow and causing application termination.

cpp

 }
 
 void tupleConstruct(Stack& stack, size_t num_inputs) {
+  if (num_inputs > stack.size()) {
+    TORCH_CHECK(false, "Invalid number of inputs: ", num_inputs);
+  }
   switch (num_inputs) {
     case 0:
       stack.emplace_back(c10::ivalue::Tuple::create());

Source: GitHub PyTorch Commit

The patch adds explicit bounds checking to validate that num_inputs does not exceed stack.size() before proceeding with the tuple construction operation.

Detection Methods for CVE-2024-31580

Indicators of Compromise

Unexpected crashes or segmentation faults in PyTorch-based applications
Application logs showing heap-related memory errors or buffer overflows
Abnormal termination of machine learning inference services without clear cause
Core dumps indicating memory access violations in vararg_functions.cpp

Detection Strategies

Monitor application stability for PyTorch-dependent services experiencing unexpected crashes
Implement memory sanitizer tools (AddressSanitizer, Valgrind) in development and staging environments to detect heap overflow conditions
Review application logs for TORCH_CHECK failures or memory access exceptions
Deploy runtime application self-protection (RASP) solutions that can detect memory corruption attempts

Monitoring Recommendations

Enable heap buffer overflow detection through compiler instrumentation (-fsanitize=address) during testing
Set up alerting for repeated PyTorch application crashes in production environments
Monitor system stability metrics for services running vulnerable PyTorch versions
Implement application health checks that can detect denial of service conditions

How to Mitigate CVE-2024-31580

Immediate Actions Required

Upgrade PyTorch to version v2.2.0 or later immediately
Audit all systems and containers running PyTorch to identify vulnerable versions
Review deployment pipelines to ensure patched versions are being deployed
Implement input validation for any external data processed by PyTorch applications

Patch Information

The Linux Foundation has addressed this vulnerability in PyTorch v2.2.0. The fix adds explicit bounds checking to validate the number of inputs before tuple construction operations. The security patch is available through the official PyTorch commit.

Organizations should update their PyTorch installations using standard package management:

Workarounds

Restrict local access to systems running PyTorch applications to trusted users only
Implement input sanitization and validation for any externally-sourced data processed by PyTorch
Deploy containerized PyTorch applications with restricted memory access permissions
Consider running PyTorch workloads in sandboxed environments to limit the impact of potential crashes

bash

# Upgrade PyTorch to patched version
pip install --upgrade torch>=2.2.0

# Verify installed version
python -c "import torch; print(torch.__version__)"

# For conda environments
conda update pytorch -c pytorch