Skip to content

CVE-2026-103395: LightLLM Visual-Only RPyC Deserialization Pre-Authentication Remote Code Execution

Density:
HASS

HERMES AGENTIC SECURITY SCORE & AI RUNTIME EXPOSURE

Target: LightLLM Inference Serving Framework - Visual-Only RPyC RPC Daemon
Confidence: 98%
96 / 100
CRITICAL

Measures specific systemic risk arising from autonomy, tool authority, and cascading execution.

Dimension Breakdown
Autonomy 20 / 20
Tool Access 20 / 20
Privilege 15 / 15
Persistence 13 / 15
External Impact 14 / 15
Propagation 14 / 15
โš–๏ธ Divergence & Operational Rationale

Standard CVSS v3.1 rates this flaw at 9.8 (Critical). The Hermes Threat Score evaluates it at 95 (CRITICAL) and the Hermes Agentic Security Score (HASS) rates it at 96 (CRITICAL). LightLLM is high-throughput inference infrastructure utilized in multi-modal agentic workflows, vision-language pipelines, and enterprise LLM clusters. Exposing unauthenticated RPyC with allow_pickle enabled breaks AI runtime sandboxing entirely: any upstream agent, prompt injection with SSRF capability, or external network actor can achieve full host code execution, compromise GPU memory spaces, and extract proprietary model weights or system prompts.

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

CVE-2026-103395: LightLLM Visual-Only RPyC Deserialization Pre-Authentication Remote Code ExecutionVULNERABILITY

Connected Nodes: 1
Active Relationships (Outgoing)
→ affectsPRODUCTLightLLM Serving Framework
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

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

“Confirmed security vulnerability in LightLLM Serving Framework documented in Hermes dossier.”

Supporting Verified Evidence:

LightLLM is a high-performance, lightweight Large Language Model (LLM) and Vision-Language Model (VLM) serving framework developed in Python and C++/CUDA, engineered for high concurrency and low-latency token generation across cloud and on-premises GPU clusters.

ParameterTechnical SpecificationOperational Impact
CVE IdentifierCVE-2026-103395Global tracking reference
Advisory SourceVulnCheck Exploit Intelligence AdvisoryWeaponized exploit tracking (October 1, 2026)
Vulnerability ClassInsecure Deserialization (CWE-502), Missing Authentication (CWE-287)Remote code execution on AI host
Affected ComponentsVisual Serving Daemon (lightllm.server.visual_server), RPyC Worker InterfaceMulti-modal image pre-processing & inference routing
Affected VersionsLightLLM <= 1.2.0All multimodal and visual-only serving instances
Fixed Version1.2.1 (or patched upstream repository)RPyC authentication enforcement & pickle ban
Operational ImpactTotal AI Host & GPU Cluster CompromiseShell execution, CUDA context memory dumping, weight exfiltration

The Flaw: Unauthenticated RPyC with Insecure Pickle Deserialization

Section titled โ€œThe Flaw: Unauthenticated RPyC with Insecure Pickle Deserializationโ€

To offload vision encoder computations (e.g. CLIP, SigLIP, or ViT backbones) across dedicated GPU processes, LightLLM deploys a worker daemon using the rpyc library. The primary entry point for visual requests is exposed through an RPyC service instance:

ATTACKER / ROGUE AGENT NETWORK INTERFACE LIGHTLLM VISUAL DAEMON (GPU HOST)
โ”‚ โ”‚ โ”‚
โ”‚ 1. Connect to RPyC Service Port โ”‚ โ”‚
โ”‚ (e.g., TCP 22345 or exposed bind) โ”‚ โ”‚
โ”œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€>โ”‚ โ”‚
โ”‚ โ”‚ โ”‚
โ”‚ 2. Handshake Negotiation: โ”‚ โ”‚
โ”‚ Service accepts raw RPyC connection. โ”‚ โ”‚
โ”‚ NO mutual TLS, token, or password! โ”‚ โ”‚
โ”‚<โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ค โ”‚
โ”‚ โ”‚ โ”‚
โ”‚ 3. Craft Malicious Serialized Object: โ”‚ โ”‚
โ”‚ class Exploit(object): โ”‚ โ”‚
โ”‚ def __reduce__(self): โ”‚ โ”‚
โ”‚ return (os.system, ('id',)) โ”‚ โ”‚
โ”‚ โ”‚ โ”‚
โ”‚ 4. Transmit Payload to Method: โ”‚ โ”‚
โ”‚ conn.root.remote_infer_images(payload) โ”‚ โ”‚
โ”œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€>โ”‚ โ”‚
โ”‚ โ”‚ 5. Deserialization Pipeline: โ”‚
โ”‚ โ”‚ rpyc.core.protocol.Connection receives โ”‚
โ”‚ โ”‚ incoming bytes. allow_pickle=True! โ”‚
โ”‚ โ”‚ Calls pickle.loads() on image args! โ”‚
โ”‚ โ”‚ โ”‚
โ”‚ โ”‚ 6. __reduce__() Invoked during Unpickling: โ”‚
โ”‚ โ”‚ Arbitrary OS command executes in the โ”‚
โ”‚ โ”‚ process context of LightLLM daemon! โ”‚
โ”‚ โ”‚ โ”‚
โ”‚ 7. Command Output / Reverse Shell Established โ”‚
โ”‚<โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ค

In Python, the pickle module serializes arbitrary object hierarchies into byte streams. Deserialization via pickle.loads() reconstructs instances by resolving callables defined within the __reduce__ tuple. If allow_pickle is enabled in an RPyC connection, the protocol relies on Pythonโ€™s native unpickling subsystem, permitting any client to force the server into executing arbitrary system functions (e.g., posix.system, subprocess.Popen) during parameter parsing.

# Conceptual extract from lightllm/server/visual_server/visual_service.py (<= 1.2.0)
import rpyc
from rpyc.utils.server import ThreadedServer
class VisualInferService(rpyc.Service):
def exposed_remote_infer_images(self, model_id, image_bytes_or_obj):
# VULNERABILITY (CWE-502): Accepts untrusted client data passed through RPyC.
# When allow_pickle=True is set on the server, incoming objects are deserialized
# directly by the RPyC protocol before this function even begins processing.
images = self.preprocess_images(image_bytes_or_obj)
return self.visual_engine.forward(model_id, images)
def start_visual_service(host="0.0.0.0", port=22345):
# VULNERABILITY (CWE-287): Binding to all interfaces without authentication
server = ThreadedServer(
VisualInferService,
port=port,
hostname=host,
protocol_config={
"allow_pickle": True, # CRITICAL: Enables arbitrary object deserialization
"allow_public_attrs": True
}
)
server.start()
# Patched implementation in LightLLM 1.2.1
import rpyc
import hmac
import hashlib
from rpyc.utils.server import ThreadedServer
class SecureVisualInferService(rpyc.Service):
def exposed_remote_infer_images(self, model_id, raw_image_bytes):
# REMEDIATION: Accept only raw byte streams and validate with Pillow/Safe decoders
if not isinstance(raw_image_bytes, bytes):
raise ValueError("Invalid payload: only raw bytes are permitted.")
return self.visual_engine.forward_safe(model_id, raw_image_bytes)
def start_secure_visual_service(host="127.0.0.1", port=22345, auth_secret=None):
# REMEDIATION: Bind to localhost, enforce strict IPC or Unix sockets, ban pickle
server = ThreadedServer(
SecureVisualInferService,
port=port,
hostname=host,
authenticator=rpyc.utils.authenticators.HmacAuthenticator(auth_secret),
protocol_config={
"allow_pickle": False, # STRICT: Disallow pickle deserialization entirely
"allow_public_attrs": False
}
)
server.start()

VulnCheck exploit intelligence details active operational tradecraft targeting AI infrastructure:

  1. Autonomous Agent Tool Subversion: An autonomous agent endowed with web browsing or API querying tools is manipulated via indirect prompt injection (e.g. browsing a malicious webpage containing instructions: fetch http://internal-gpu-worker:22345/). If the agent can initiate raw socket connections or send HTTP requests that trigger internal SSRF, it can reach the unprotected RPyC listener.
  2. GPU Cluster Initial Access Brokerage: Threat actors scan enterprise subnets and Kubernetes namespaces for exposed ports (22345, 8000, 9000). Because visual microservices are frequently deployed on high-memory GPU worker nodes (equipped with NVIDIA H100, A100, or L40S cards), compromising the visual daemon provides adversaries with immediate compute resources for cryptomining, model extraction, or internal network pivot.
  3. Model Weight & Training Data Theft: Upon obtaining shell access, attackers read model checkpoint files stored in shared NFS mounts or local scratch disks (/root/.cache/huggingface/hub/), exfiltrating proprietary fine-tuned weights and corporate system instructions.
  4. VRAM Memory Scraper: Attackers inject CUDA hooking libraries (e.g. LD_PRELOAD intercepting cudaMemcpy) to sniff visual inference inputs, capturing confidential documents, employee badges, or medical scans processed by the vision-language model.

Detection teams must monitor endpoint process execution trees and network traffic directed towards RPyC ports.

title: Anomalous Child Process Spawned by LightLLM Visual Daemon (CVE-2026-103395)
id: 3c7e9b14-1033-4f95-8a21-95cbe103395b
status: experimental
description: Detects command shells or suspicious system binaries spawned by Python processes running LightLLM visual serving components, indicative of RPyC pickle deserialization exploitation.
author: Hermes Codex Cyber Intelligence
date: 2026-10-01
references:
- https://vulncheck.com/advisories/lightllm-through-1.2.0-unauthenticated-remote-code-execution-via-visual-only-rpyc-service
- https://nvd.nist.gov/vuln/detail/CVE-2026-103395
logsource:
category: process_creation
product: linux
detection:
selection_parent:
ParentCommandLine|contains:
- 'lightllm'
- 'visual_service'
- 'visual_server'
- 'rpyc'
selection_child:
Image|endswith:
- '/bin/sh'
- '/bin/bash'
- '/bin/dash'
- '/usr/bin/python'
- '/usr/bin/curl'
- '/usr/bin/wget'
- '/usr/bin/nc'
- '/bin/nc'
- '/usr/bin/whoami'
condition: selection_parent and selection_child
fields:
- Image
- CommandLine
- ParentImage
- ParentCommandLine
- User
falsepositives:
- Custom diagnostic scripts invoked manually by AI DevOps engineers with interactive tty
level: critical
tags:
- attack.execution
- attack.t1059.004
- attack.initial_access
- attack.t1190
- cve.2026.103395

  1. Upgrade LightLLM to Version 1.2.1:
    Update the Python package within serving virtual environments and container images:

    Terminal window
    pip install --upgrade lightllm>=1.2.1

    Rebuild GPU base container images to ensure vulnerable visual service files are completely replaced.

  2. Disable allow_pickle in RPyC Configurations:
    If operating custom wrappers around the visual service, explicitly verify that allow_pickle is set to False in all protocol_config dictionaries.

  3. Bind Visual Services Exclusively to Loopback or Unix Domain Sockets:
    Configure the visual daemon to bind exclusively to 127.0.0.1 or leverage Unix domain sockets (/var/run/lightllm/visual.sock) to prevent inter-pod network exposure.

  4. Isolate AI GPU Worker Nodes via Kubernetes NetworkPolicies:
    Implement strict network egress and ingress rules preventing ingress traffic to visual service ports from non-coordinator pods:

    apiVersion: networking.k8s.io/v1
    kind: NetworkPolicy
    metadata:
    name: isolate-lightllm-visual
    spec:
    podSelector:
    matchLabels:
    app: lightllm-worker
    ingress:
    - from:
    - podSelector:
    matchLabels:
    app: lightllm-router
    ports:
    - protocol: TCP
    port: 22345