Critical Vulnerability Research

vLLM RCE: Send a Malicious Video, Take Over AI Servers

CVE-2026-22778 allows attackers to achieve Remote Code Execution on vLLM servers by sending a specially crafted video URL. No authentication required.Our stealth reconnaissance found 175,000+ exposed LLM servers across 130 countries.

Sable Security Research Team
February 4, 2026
0 views
10 min read

TL;DR: What We Found

175K+
Servers Exposed
globally on Shodan
130
Countries
affected worldwide
91,403
Attack Sessions
Oct 2025 - Jan 2026
9.8
CVSS Score
CRITICAL severity

Impact: Complete server compromise without authentication. Affects vLLM 0.8.3 - 0.14.0 with multimodal video support. Patch available in 0.14.1+.

The Attack Chain: 2-Stage Exploitation

The vulnerability exploits a heap overflow in OpenCV's JPEG2000 decoder, combined with an information leak that bypasses ASLR. Together, they provide reliable RCE.

1Information Leak (ASLR Bypass)
Attacker sends invalid/malformed image to multimodal endpoint
PIL (Pillow) throws exception with heap memory address
vLLM returns error to client with BytesIO object address
ASLR entropy reduced from ~4 billion to ~8 attempts
2Heap Overflow → RCE
Attacker sends malicious video URL with corrupted JPEG2000
OpenCV/FFmpeg (v5.1.x) processes video frames
Manipulated "cdef" box redirects Y-plane to U-buffer
9,600 bytes overflow → overwrites function pointers → system()
heap-overflow.md
# JPEG2000 Channel Overflow Calculation
Y-plane data: 9,600 bytes
U-buffer size: 2,400 bytes
Overflow: 7,200 bytes → RCE

Vulnerable Models (Video-Capable)

LLaVA
LLaVA-NeXTLLaVA-OneVision
Qwen-VL
Qwen2-VLQwen3-VL
InternVL
InternLM-XComposer
Phi-Vision
Phi-3-VisionPhi-3.5-Vision
Others
PixtralMolmoVideo-LLaMA
Vulnerable:0.8.3 - 0.14.0
Patched:≥ 0.14.1

Global Exposure: 175,000+ Servers

Geographic Distribution

🇨🇳China
30%
🇺🇸United States
15%
🇩🇪Germany
8%
🇫🇷France
6%
🇰🇷South Korea
5%
🇮🇳India
4%
🇷🇺Russia
4%
🇸🇬Singapore
3%
🇧🇷Brazil
3%
🌍Others
22%

Exposure Statistics

OpenAI-compatible API
Same attack vector
88.9%
Tool-calling enabled
Higher risk
48%
No safety guardrails
Completely open
201
Est. abuse cost/day
Per compromised server
$46K

Active Attack Campaign (GreyNoise Data)

Oct 2025 - Jan 2026
91,403
Total attack sessions
Dec 28 - Jan 8 (11 days)
80,469
88% of all attacks

88% acceleration: Most attacks occurred in just 11 days, indicating active exploitation campaigns.

Our Stealth Reconnaissance

We conducted stealth reconnaissance through Tor to identify exposed vLLM endpoints. All scanning was passive - no exploitation attempts were made.

ProviderStatusModelsMultimodalRisk
DeepInfraOPEN1326Medium
OpenRouterOPENRouterN/ALow
Together AIAUTHUnknownYesLow
Responsible Reconnaissance

All scans routed through Tor (exit: 205.185.113.8). No exploitation attempts. Detection and documentation only.

LLMjacking: The Business Impact

Compromised vLLM servers are monetized through "LLMjacking" - unauthorized use of AI infrastructure for spam, disinformation, crypto mining, or API resale.

Compute Theft

GPU resources used for unauthorized inference or crypto mining

Data Exfiltration

Access to models, training data, and conversation history

API Resale

Stolen access sold on criminal marketplaces (silver.inc)

Lateral Movement

Pivot to other systems in the network

Operation Bizarre Bazaar

Active criminal marketplace (silver.inc) selling access to compromised LLM endpoints. Documented by Trend Micro and Pillar Security.

Immediate Mitigation Steps

P024hUpdate vLLM
pip install --upgrade vllm>=0.14.1
P024hDisable Video Models
# Use text-only models if video not needed
# Remove: llava, qwen-vl, phi-vision, etc.
P11 weekEnable Authentication
vllm serve model_name --api-key YOUR_SECRET_KEY
P11 weekFirewall Rules
iptables -A INPUT -p tcp --dport 8000 -s TRUSTED_IP -j ACCEPT
iptables -A INPUT -p tcp --dport 8000 -j DROP
P21 monthReverse Proxy + WAF
# nginx.conf
location /v1/ {
  auth_basic "vLLM API";
  proxy_pass http://localhost:8000;
}

Disclosure Timeline

Jan 2026
CVE-2026-22778 discovered by Orca Security
Jan 2026
CVE officially assigned (CVSS 9.8)
Jan 2026
vLLM 0.14.1 released with patch
Jan 29, 2026
SentinelOne/Censys publish exposure statistics
Feb 4, 2026
Sable Security completes stealth reconnaissance

Running vLLM in Production?

Get a professional security assessment to identify vulnerabilities before attackers do. We specialize in AI infrastructure security.

This research was conducted for defensive purposes only. All testing was stealth reconnaissance. No user data was accessed or stored.