CVE-2025-65891
📋 TL;DR
A GPU device-ID validation flaw in OneFlow v0.9.0 allows attackers to trigger a Denial of Service (DoS) by calling flow.cuda.get_device_properties() with an invalid or negative device index. This affects systems running OneFlow v0.9.0 with CUDA GPU support enabled. The vulnerability can crash the application or cause GPU-related instability.
💻 Affected Systems
- OneFlow
📦 What is this software?
Oneflow by Oneflow
⚠️ Risk & Real-World Impact
Worst Case
Complete application crash leading to service disruption, potential data loss in active processing, and GPU driver instability requiring system reboot.
Likely Case
Application crash or hang when malicious input triggers the vulnerability, requiring process restart and interrupting GPU-accelerated workloads.
If Mitigated
Minimal impact if proper input validation is implemented or vulnerable function calls are restricted.
🎯 Exploit Status
Exploitation requires ability to call the vulnerable function with malicious input, typically through code execution or API access.
🛠️ Fix & Mitigation
✅ Official Fix
Patch Version: v0.9.1 or later
Vendor Advisory: https://github.com/Oneflow-Inc/oneflow/issues/10661
Restart Required: Yes
Instructions:
1. Check current version with 'pip show oneflow'. 2. Upgrade using 'pip install --upgrade oneflow'. 3. Restart all OneFlow applications and services.
🔧 Temporary Workarounds
Input Validation Wrapper
allWrap calls to flow.cuda.get_device_properties() with validation to ensure device index is non-negative and within valid range.
# Python wrapper example:
def safe_get_device_properties(device_id):
if device_id < 0:
raise ValueError('Invalid device ID')
return flow.cuda.get_device_properties(device_id)
Disable GPU Support
linuxRun OneFlow in CPU-only mode if GPU acceleration is not required.
export CUDA_VISIBLE_DEVICES=''
or set environment variable before running OneFlow applications
🧯 If You Can't Patch
- Implement strict input validation for all user-controlled inputs that could reach the vulnerable function.
- Monitor application logs for crashes or errors related to flow.cuda.get_device_properties() calls and implement rate limiting.
🔍 How to Verify
Check if Vulnerable:
Check if running OneFlow v0.9.0 with 'pip show oneflow' and verify CUDA is available with 'python -c "import oneflow; print(oneflow.cuda.is_available())"'.Check Version:
pip show oneflow | grep VersionVerify Fix Applied:
After upgrading, test with 'python -c "import oneflow; oneflow.cuda.get_device_properties(-1)"' which should raise a proper error instead of crashing.📡 Detection & Monitoring
Log Indicators:
- Application crashes with GPU-related errors
- Stack traces containing 'flow.cuda.get_device_properties'
- Unexpected process termination
Network Indicators:
- Sudden drop in ML inference/training throughput
- Unusual API calls to GPU-related endpoints
SIEM Query:
source='application.logs' AND ("flow.cuda.get_device_properties" OR "GPU error" OR "cuda illegal memory access")