CVE-2025-66455

Summary

LMDeploy's PyTorch DistServe/PD-disaggregation control plane used

recv_pyobj() to deserialize messages received through a ZeroMQ PULL

socket. PyZMQ implements recv_pyobj() using Python pickle

deserialization, which can execute arbitrary code while reconstructing

an object.

The peer address used by the receiver was supplied through the

POST /distserve/p2p_connect HTTP endpoint. An attacker who could reach

an affected DistServe API server could cause the server to connect to an

attacker-controlled ZeroMQ endpoint and deserialize a crafted pickle

payload.

API-key authentication is not enabled unless the operator explicitly

configures it. As a result, affected DistServe deployments without API

keys allowed unauthenticated remote code execution with the privileges

of the LMDeploy serving process.

This issue affects the PyTorch backend when PD-disaggregation/DistServe

is enabled. Ordinary deployments that do not use the affected

disaggregated-serving path do not expose this data flow.

Affected components

• HTTP entry point:

  lmdeploy/serve/openai/endpoints/distserve.py,

  POST /distserve/p2p_connect

• Attacker-controlled peer address:

  DistServeConnectionRequest.remoteengineendpointinfo.zmqaddress

• Vulnerable receiver:

  lmdeploy/pytorch/disagg/conn/engine_conn.py,

  EngineP2PConnection.handlezmqrecv()

• Unsafe operation: recv_pyobj(), which performs pickle deserialization

Vulnerable data flow

1. A caller submits a DistServe P2P connection request containing a

   ZeroMQ address.

1. The LMDeploy engine connects its ZeroMQ PULL socket to that address.
2. handlezmqrecv() receives messages using recv_pyobj().
3. A malicious peer sends a crafted pickle object.
4. Python code executes during deserialization, before LMDeploy can

   perform any type or field validation.

  A type check performed after recv_pyobj() cannot mitigate this issue

  because pickle payload execution occurs during deserialization.

Impact

Successful exploitation allows arbitrary code execution as the

LMDeploy serving process. This can expose model weights, prompts,

credentials, attached storage, cluster-network services, and host or

GPU resources. An attacker may also modify or terminate the serving

process.

Affected versions

Affected versions:

• lmdeploy >= 0.9.2, < 0.16.0

The vulnerable P2P receiver was introduced in commit b0b705f7.

Remediation

The issue was fixed by replacing the pickle-based ZeroMQ protocol with

JSON serialization:

• send_pyobj() was replaced with send_json().
• recv_pyobj() was replaced with recv_json().
• Received objects are validated using the

  DistServeCacheFreeRequest Pydantic schema before use.

• Invalid or off-schema messages are rejected without terminating the

  receive loop.

Fix commit:

https://github.com/InternLM/lmdeploy/commit/f05b4ad8bf2e2d84101a1d63b3c44fadd99223b2

The fix was released in LMDeploy 0.16.0.

Workarounds

Users who cannot upgrade immediately should:

• Prevent untrusted clients from reaching /distserve/* endpoints.
• Restrict the DistServe HTTP and ZeroMQ control planes to trusted

  cluster networks.

• Configure API-key authentication.
• Block arbitrary outbound ZeroMQ connections from serving nodes.

These measures reduce exposure but do not make pickle deserialization

safe. Upgrading to LMDeploy 0.16.0 or later is recommended.