CVE-2025-70999
📋 TL;DR
A GPU device-ID validation flaw in OneFlow's CUDA component allows attackers to trigger a Denial of Service (DoS) by providing a crafted device ID. This affects OneFlow v0.9.0 users who utilize GPU acceleration. The vulnerability can crash the application when processing malicious input.
💻 Affected Systems
- OneFlow
📦 What is this software?
Oneflow by Oneflow
⚠️ Risk & Real-World Impact
Worst Case
Complete service disruption where the OneFlow application crashes and becomes unavailable, potentially affecting dependent services or workflows.
Likely Case
Application crash or hang when processing malicious device IDs, requiring manual restart and causing temporary service interruption.
If Mitigated
Minimal impact if input validation is implemented externally or if the vulnerable function is not exposed to untrusted sources.
🎯 Exploit Status
Exploitation requires access to call flow.cuda.get_device_capability() with malicious input. This could be through user-controlled parameters, API calls, or data processing pipelines.
🛠️ Fix & Mitigation
✅ Official Fix
Patch Version: Check GitHub issue #10660 for latest patched version
Vendor Advisory: https://github.com/Oneflow-Inc/oneflow/issues/10660
Restart Required: Yes
Instructions:
1. Check the GitHub issue #10660 for patch availability
2. Update OneFlow to the latest patched version
3. Restart any running OneFlow applications or services
🔧 Temporary Workarounds
Input Validation Wrapper
allImplement input validation before calling flow.cuda.get_device_capability() to ensure device IDs are within valid range.
# Python example wrapper
import oneflow as flow
def safe_get_device_capability(device_id):
if device_id < 0 or device_id >= flow.cuda.device_count():
raise ValueError('Invalid device ID')
return flow.cuda.get_device_capability(device_id)
Disable GPU Support
allRun OneFlow in CPU-only mode if GPU acceleration is not required.
# Set environment variable to disable CUDA
import os
os.environ['CUDA_VISIBLE_DEVICES'] = ''
🧯 If You Can't Patch
- Implement strict input validation for all device ID parameters before passing to CUDA functions
- Isolate OneFlow services in containers or VMs to limit blast radius if DoS occurs
🔍 How to Verify
Check if Vulnerable:
Check if running OneFlow v0.9.0 with CUDA support: import oneflow; print(oneflow.__version__); print(oneflow.cuda.is_available())Check Version:
python -c "import oneflow; print(oneflow.__version__)"Verify Fix Applied:
After updating, verify version is newer than v0.9.0 and test with invalid device IDs to ensure proper error handling.📡 Detection & Monitoring
Log Indicators:
- Application crashes or segmentation faults related to CUDA/GPU operations
- Error messages mentioning device ID validation or out-of-bounds access
Network Indicators:
- Unusual patterns of requests to GPU-related API endpoints
- Sudden service unavailability after device ID parameter changes
SIEM Query:
source='application.logs' AND ("segmentation fault" OR "CUDA error" OR "device ID")