/

CVE-2026-90919

Severity:

CRITICAL RISK

(9.8)

Unsafe Deserialization (CWE-502) in LightLLM Config Server

CVE-2026-90919

Deferred

Vulnerability type

Category:

Unsafe Deserialization

CWE:

CWE-502

Impact:

Remote Code Execution

Severity:

CRITICAL RISK

(9.8)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

LightLLM's Config Server unsafely deserializes data received through an unauthenticated WebSocket endpoint. An attacker who can reach the Config Server port may execute arbitrary code with the privileges of the server process.

Affected context

The vulnerability occurs in LightLLM deployments using the Config Server through version 1.2.0, where the unauthenticated /visual_register WebSocket endpoint processes the first client frame as a serialized Python object. Risk is highest when the Config Server port is reachable by untrusted or insufficiently restricted networks.

Exploit preconditions

  • Attack surface: The Config Server's /visual_register WebSocket endpoint - Required access: Network connectivity to the Config Server port; endpoint

authentication is not required

  • Trigger point: Processing of the first client WebSocket frame - Why exploitation is possible: The received data is passed to an unsafe

deserialization routine that can reconstruct attacker-controlled objects

and invoke their associated behavior

Why this happens

The underlying bug is the use of pickle.loads() on data received from an unauthenticated client. Python pickle deserialization is not a data-only parsing mechanism: reconstructing objects can invoke attacker-influenced behavior. This maps to CWE-502, Unsafe Deserialization, because untrusted serialized input is treated as trusted executable object data.

Potential impact

  • An attacker may execute arbitrary code in the Config Server process.
  • The attacker could access or alter data available to that process.
  • The service may be used to affect connected workers or other internal systems.
  • Compromise could enable persistence, lateral movement, or service disruption.

General mitigation guidance

Mitigation generally centers on removing unsafe deserialization from unauthenticated data paths. Secure-by-default designs can use a restricted, data-only serialization format with explicit schemas and validation, while treating all network-supplied frames as untrusted. Access to the Config Server should be limited to trusted clients through authentication, authorization, network segmentation, and narrowly scoped exposure. Running the service with least privilege, restricted filesystem and network permissions, and isolated credentials can reduce the consequences of a compromise.

How teams detect issues like this

Detection can include static analysis for pickle.loads(), related pickle APIs, and deserialization of network-controlled data. Code review should examine unauthenticated WebSocket handlers, dynamic object reconstruction, and validation performed after deserialization. Runtime monitoring can flag unexpected connections to the Config Server, anomalous WebSocket traffic, child-process creation, and behavior inconsistent with the service role. CI and security policy checks may enforce approved data-only serialization formats and prohibit unsafe deserialization in network-facing components.

Sources

SourceLink
github.comgithub.com/ModelTC/LightLLM/blob/v1.2.0/lightllm/server/config_server/api_http.py#L79-L86
github.comgithub.com/ModelTC/LightLLM/issues/1563
github.comgithub.com/ModelTC/lightllm
vulncheck.comvulncheck.com/advisories/lightllm-through-1.2.0-unauthenticated-remote-code-execution-via-config-server-pickle-deserialization

Frequently asked

questions

Yes. An attacker generally only needs network reachability to the Config Server port because the affected WebSocket endpoint does not require authentication.

The attacker must be able to connect to the Config Server and send data to the affected endpoint. Exposure depends on deployment networking, firewall rules, and service placement.

Do not deserialize untrusted network input with Python pickle. Use an authenticated, data-only format with strict schema validation and explicit permitted data types.

Apply secure serialization policies, review all network-facing parsers, restrict service exposure, run components with least privilege, and use automated checks for unsafe deserialization APIs.

Find out what's already
exploitable in your codebase.