Skip to content

CVE-2026-108719: LLMGateway Blind SSRF via Video-Generation callback_url Webhook Handling

Density:
C-Level Executive Briefing CVE-2026-108719 CRITICAL SCORE 92/100

Critical blind SSRF vulnerability in LLMGateway asynchronous video generation callback loop permitting unauthenticated cloud metadata extraction and internal network pivoting.

AGENTIC

HERMES AGENTIC SECURITY SCORE & AI REVERSE PROXY EVALUATION

Target: LLMGateway โ€” Asynchronous Video Generation Webhook Dispatcher
Confidence: 97%
89 / 100
CRITICAL

Measures real-world operational relevance, exploit weaponization, and active threat posture.

Dimension Breakdown
Autonomy 16 / 20
Tool Access 18 / 20
Privilege 14 / 20
Persistence 11 / 20
External Impact 15 / 20
Propagation 15 / 20
โš–๏ธ Divergence & Operational Rationale

CVSS v3.1 rates this vulnerability at 9.3 (CRITICAL) due to the scope change (S:C) allowing access to cloud instance metadata and private cluster services. The Hermes Agentic Security Score (HASS) rates this at 89 (CRITICAL). As an API gateway deployed at the perimeter to broker requests to multiple generative AI and LLM models, LLMGateway frequently runs with elevated cloud permissions to access GPU nodes, cloud storage buckets, and model endpoints. A blind SSRF flaw in its asynchronous callback loop allows attackers to retrieve temporary cloud instance profile credentials (AWS IMDSv1/v2, GCP metadata) and pivot directly into cloud control planes.

๐Ÿ•ธ๏ธ Connected Knowledge Graph & Provenance

CVE-2026-108719: LLMGateway Blind SSRF via Video-Generation callback_url Webhook HandlingVULNERABILITY

Connected Nodes: 1
Active Relationships (Outgoing)
→ affectsPRODUCTLLMGateway Multimodal AI Reverse Proxy
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

๐Ÿ” Why is this related? (Evidence & Provenance)

“Confirmed security vulnerability in LLMGateway Multimodal AI Reverse Proxy documented in Hermes dossier.”

Supporting Verified Evidence:

LLMGateway sits between external clients and upstream foundation model APIs, providing load balancing, caching, and rate limiting.

ParameterTechnical SpecificationOperational Details
CVE IdentifierCVE-2026-108719Canonical vulnerability tracking identifier
Affected ProductLLMGatewayMultimodal AI reverse proxy and perimeter router
Vulnerable Componentgateway/callbacks/dispatcher.pyAsynchronous webhook delivery service
Impacted VersionsAll versions through 1.20.0Cloud, Kubernetes, and bare-metal AI gateway deployments
Corrective Release1.20.1Strict egress policy validation and header sanitization
Vulnerability ClassSSRF (CWE-918, CWE-610)Unrestricted outbound HTTP request dispatch
Execution ContextPython / FastAPI / Docker / KubernetesAI perimeter gateway pods with cloud IAM role attachments

Vulnerable Code Path: Missing Egress Validation on Webhook Dispatch

Section titled โ€œVulnerable Code Path: Missing Egress Validation on Webhook Dispatchโ€

In gateway/callbacks/dispatcher.py, the callback execution routine retrieves the target URL from the incoming request payload and executes an HTTP POST request without verifying if the target hostname resolves to loopback, RFC 1918, or link-local addresses:

gateway/callbacks/dispatcher.py
import httpx
async def dispatch_completion_webhook(task_id: str, callback_url: str, result_payload: dict, client_headers: dict):
# FLAW 1: No URL validation or IP resolution check
# FLAW 2: Forwards client and internal headers to the unvalidated callback target!
async with httpx.AsyncClient(timeout=10.0) as client:
response = await client.post(
callback_url,
json={"task_id": task_id, "status": "completed", "output": result_payload},
headers={"Authorization": client_headers.get("Authorization"), "X-Gateway-Node": "ai-proxy-prod"}
)
return response.status_code
Attacker Request LLMGateway Perimeter Pod Cloud Metadata / Internal Pods
+-------------------------+ +-------------------------------+ +-------------------------------+
| POST /v1/video/ | | Video Generation Task Created | | 169.254.169.254 (IMDSv2/v1) |
| generations | ------> | | | internal-etcd.cluster.local |
| callback_url: | | Asynchronous Dispatcher fires | ---> | kube-apiserver.default.svc |
| http://169.254.169.254/ | | HTTP POST from Pod Network | +-------------------------------+
| latest/meta-data/ | +-------------------------------+ |
+-------------------------+ v
+-------------------------------+
| Internal Service Disrupted |
| or Metadata Probed by Attacker|
+-------------------------------+

  1. Payload Submission: An unauthenticated or low-privilege attacker sends a JSON payload to /v1/video/generations:
    {
    "prompt": "Render 3D corporate logo rotation",
    "model": "video-gen-v2",
    "callback_url": "http://169.254.169.254/latest/meta-data/iam/security-credentials/"
    }
  2. Task Processing & Callback Trigger: The gateway queues the request. Upon generation completion (or premature cancellation), the background dispatcher initiates an outbound HTTP request to the target metadata URI.
  3. Internal Network Scanning & Cloud Pivot: By manipulating target paths and monitoring gateway error metrics or response delays, attackers map internal network services, trigger administrative actions on vulnerable internal microservices, or leverage secondary DNS-rebinding listeners to capture exfiltrated authentication headers.

title: LLMGateway Asynchronous Callback SSRF Probe
id: c418a920-8114-4a21-9920-202610871901
status: experimental
description: Detects outbound network traffic from AI gateway container pods directed toward internal IP ranges or cloud metadata endpoints.
references:
- https://vulncheck.com/advisories/llmgateway-through-1.20.0-blind-ssrf-via-video-generation-callback-url
author: Hermes Codex CTI Team
date: 2026/10/12
logsource:
category: network_connection
product: linux
detection:
selection_process:
Image|endswith:
- '/python3'
- '/uvicorn'
- '/llmgateway'
selection_dest:
DestinationIp:
- '169.254.169.254'
- '127.0.0.1'
DestinationIp|startswith:
- '10.'
- '172.16.'
- '172.17.'
- '172.18.'
- '192.168.'
condition: selection_process and selection_dest
fields:
- Image
- CommandLine
- DestinationIp
- DestinationPort
level: critical
tags:
- attack.initial_access
- attack.t1190

Upgrade LLMGateway to version 1.20.1 or later:

Terminal window
pip install --upgrade llmgateway>=1.20.1
# Or update the container image:
docker pull ghcr.io/llmgateway/llmgateway:1.20.1

Enforce Kubernetes NetworkPolicies to block egress to cloud metadata from AI reverse proxy pods:

apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
name: block-metadata-egress
namespace: ai-gateway
spec:
podSelector:
matchLabels:
app: llmgateway
egress:
- to:
- ipBlock:
cidr: 0.0.0.0/0
except:
- 169.254.169.254/32
- 10.0.0.0/8
- 172.16.0.0/12
- 192.168.0.0/16
policyTypes:
- Egress