CVE-2025-23321
📋 TL;DR
NVIDIA Triton Inference Server contains a divide-by-zero vulnerability in request processing that could cause denial of service. Attackers can exploit this by sending specially crafted invalid requests to crash the server. This affects all deployments using vulnerable versions of NVIDIA Triton Inference Server.
💻 Affected Systems
- NVIDIA Triton Inference Server
📦 What is this software?
⚠️ Risk & Real-World Impact
Worst Case
Complete service disruption requiring manual restart of Triton Inference Server, potentially affecting all inference workloads and dependent applications.
Likely Case
Temporary denial of service affecting specific inference endpoints until service restart, with potential data loss for in-flight requests.
If Mitigated
Minimal impact with automatic restart mechanisms and request validation preventing exploitation.
🎯 Exploit Status
Exploitation requires sending malformed requests but no authentication is needed.
🛠️ Fix & Mitigation
✅ Official Fix
Patch Version: 24.09 or later
Vendor Advisory: https://nvidia.custhelp.com/app/answers/detail/a_id/5687
Restart Required: Yes
Instructions:
1. Download Triton Inference Server version 24.09 or later from NVIDIA NGC. 2. Stop the current Triton service. 3. Install the updated version. 4. Restart the Triton service.
🔧 Temporary Workarounds
Request Validation Proxy
allDeploy a reverse proxy or API gateway to validate and filter incoming requests before they reach Triton.
Network Segmentation
allRestrict access to Triton Inference Server to only trusted clients and networks.
🧯 If You Can't Patch
- Implement strict network access controls to limit Triton server exposure
- Deploy monitoring and automatic restart mechanisms for Triton service
🔍 How to Verify
Check if Vulnerable:
Check Triton version using 'tritonserver --version' or examine container image tags. Versions before 24.09 are vulnerable.Check Version:
tritonserver --versionVerify Fix Applied:
Confirm version is 24.09 or later and test with malformed request simulation to ensure service remains stable.📡 Detection & Monitoring
Log Indicators:
- Unexpected server crashes
- Divide-by-zero errors in logs
- Abnormal termination of Triton process
Network Indicators:
- Spike in malformed HTTP/GRPC requests to Triton endpoints
- Unusual request patterns from single sources
SIEM Query:
source="triton" AND ("divide by zero" OR "segmentation fault" OR "unexpected termination")