What is the CVSS score of EUVD-2025-210969?

EUVD-2025-210969 has a CVSS 3.1 base score of 9.8 (Critical). CVSS vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H.

Which versions of LMDeploy are affected by EUVD-2025-210969?

LMDeploy versions 0.9.2 through 0.15.x are vulnerable to unauthenticated remote code execution, allowing attackers to compromise the serving process and potentially access sensitive data or disrupt…. A patch is available.

How to fix or mitigate EUVD-2025-210969?

Within 24 hours, identify all deployments of LMDeploy versions 0.9.2 through 0.15.x, particularly those using DistServe/PD-disaggregation, and apply the vendor patch by upgrading to LMDeploy 0.16.0; if immediate patching is not possible, enable API-key authentication on the serving endpoint as a temporary mitigation. Within 7 days, verify that all affected instances have been patched and conduct a security review of exposed endpoints to ensure no unauthorized access occurred. Within 30 days, implement a vulnerability management process to track and promptly patch critical vulnerabilities in AI serving frameworks, and consider network segmentation to limit exposure of such services.

LMDeploy EUVD-2025-210969

| CVE-2025-66455 CRITICAL

Deserialization of Untrusted Data (CWE-502)

2026-09-18 https://github.com/InternLM/lmdeploy

GHSA-2vh9-42vm-xmv2

RCE Python Deserialization

9.8

CVSS 3.1 · Vendor: https://github.com/InternLM/lmdeploy

Severity by source

Vendor (https://github.com/InternLM/lmdeploy) PRIMARY

9.8 CRITICAL

AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H

vuln.today AI

9.8 CRITICAL

Network-reachable endpoint, default no auth (PR:N/UI:N), low-complexity pickle RCE granting full code execution as the serving process (C:H/I:H/A:H); scope unchanged as impact stays within the process privileges.

3.1 AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H

4.0 AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N

Primary rating from Vendor (https://github.com/InternLM/lmdeploy).

CVSS VectorVendor: https://github.com/InternLM/lmdeploy

Attack Vector

Network

Attack Complexity

Low

Privileges Required

None

User Interaction

None

Scope

Unchanged

Confidentiality

High

Integrity

High

Availability

High

Lifecycle Timeline

POC Analysis Generated

Sep 18, 2026 - 19:26 vuln.today

Analysis Generated

Sep 18, 2026 - 17:35 vuln.today

CVE Published

Sep 18, 2026 - 17:03 github-advisory

CRITICAL 9.8

DescriptionCVE.org

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.remote_engine_endpoint_info.zmq_address

Vulnerable receiver:

lmdeploy/pytorch/disagg/conn/engine_conn.py, EngineP2PConnection.handle_zmq_recv()

Unsafe operation: recv_pyobj(), which performs pickle deserialization

Vulnerable data flow

A caller submits a DistServe P2P connection request containing a

ZeroMQ address.

The LMDeploy engine connects its ZeroMQ PULL socket to that address.
handle_zmq_recv() receives messages using recv_pyobj().
A malicious peer sends a crafted pickle object.
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.

Articles & Coverage 1

News Critical RCE in LMDeploy 0.9.2-0.15.x DistServe P2P Endpoint - CVE-2025-66455 Sep 19, 2026

AnalysisAI

Unauthenticated remote code execution in LMDeploy 0.9.2 through 0.15.x lets any attacker who can reach a DistServe/PD-disaggregation deployment's POST /distserve/p2p_connect endpoint run arbitrary code with the privileges of the LMDeploy serving process. The endpoint accepts an attacker-supplied ZeroMQ peer address; the engine connects to it and deserializes the reply with PyZMQ's recv_pyobj(), which is backed by Python pickle and therefore executes code during object reconstruction, before any type or schema check can run. Rated CVSS 3.1 9.8 (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H, CWE-502); exploitation is unauthenticated only in the default configuration where the operator has not enabled API-key authentication, and no public exploit code has been identified at time of analysis. A vendor patch is available in LMDeploy 0.16.0, which replaces the pickle-based ZeroMQ protocol with JSON plus Pydantic validation.

Technical ContextAI

The flaw is CWE-502 (Deserialization of Untrusted Data) rooted in the PD-disaggregation (prefill/decode separation, branded DistServe) control plane of LMDeploy's PyTorch backend. In that architecture, separate prefill and decode engine instances communicate over ZeroMQ PUSH/PULL sockets to coordinate cache-freeing between peers. The vulnerable receiver, EngineP2PConnection.handle_zmq_recv() in lmdeploy/pytorch/disagg/conn/engine_conn.py, used PyZMQ's recv_pyobj(), which internally calls pickle.loads() on peer-supplied bytes; pickle's object reconstruction invokes arbitrary callables via __reduce__, so code executes inside the deserialization step itself, defeating any isinstance() or field check written afterwards. The peer address was neither hard-coded nor authenticated: it was taken from DistServeConnectionRequest.remote_engine_endpoint_info.zmq_address submitted to the HTTP route in lmdeploy/serve/openai/endpoints/distserve.py, so an attacker simply points the server's PULL socket at their own ZeroMQ endpoint and sends a crafted pickle. The upstream fix (commit f05b4ad8bf2e2d84101a1d63b3c44fadd99223b2) swaps send_pyobj()/recv_pyobj() for send_json()/recv_json(), validates incoming objects against the DistServeCacheFreeRequest Pydantic schema with model_validate(), and catches ValueError/ValidationError so malformed messages are logged and skipped rather than terminating the receive loop. Affected package is the pip distribution pkg:pip/lmdeploy; the vulnerable receiver was introduced in commit b0b705f7.

RemediationAI

Upgrade to LMDeploy 0.16.0 or later, which is the vendor-released fix for this issue and replaces the pickle wire format with JSON plus DistServeCacheFreeRequest Pydantic schema validation (fix commit: https://github.com/InternLM/lmdeploy/commit/f05b4ad8bf2e2d84101a1d63b3c44fadd99223b2; release: https://github.com/InternLM/lmdeploy/releases/tag/v0.16.0; advisory: https://github.com/InternLM/lmdeploy/security/advisories/GHSA-2vh9-42vm-xmv2). If an immediate upgrade is not possible, apply compensating controls knowing they narrow rather than eliminate the pickle deserialization path: enable API-key authentication on the LMDeploy server so that /distserve/* requires credentials (this restores authentication but does nothing if a key holder is malicious and does not fix the unsafe deserialization itself); restrict the DistServe HTTP endpoint and the ZeroMQ control plane to a trusted cluster network via firewall or reverse-proxy rules so untrusted clients cannot reach /distserve/p2p_connect (this may break legitimate cross-node disaggregated serving if peer ranges are not scoped correctly); and block arbitrary outbound ZeroMQ/TCP connections from serving nodes with egress policy, which limits the attacker's ability to redirect the PULL socket to their own endpoint but can disrupt legitimate peer-to-peer communication between prefill and decode engines if the allowed peer set is not enumerated. Operators who do not need PD-disaggregation should disable DistServe entirely, which removes the affected endpoint and receiver from the running service. Vendor guidance notes that upgrading remains the recommended remedy because workarounds do not make pickle deserialization safe.