CVE-2026-76850 InternLM CVE debrief

Who should care

Defenders of LMDeploy deployments, particularly those using disaggregated serving, should assess exposure and apply patches or mitigations to prevent exploitation. This involves verifying affected versions and deployments, patching or mitigating affected deployments, and monitoring for suspicious activity on ZMQ endpoints. Additionally, defenders should review and apply patches from the vendor, specifically updating to version 0.16.0 or later, disable dis

Why it matters

CVE-2026-76850 is a critical vulnerability in LMDeploy that allows for remote code execution via unsafe pickle deserialization. Defenders of LMDeploy deployments, particularly those using disaggregated serving, should assess exposure and apply patches or mitigations to prevent exploitation.

• Remote code execution is possible in deployments with disaggregated serving enabled
• Verification of affected versions and deployments is required
• Patching or mitigating affected deployments is necessary to prevent exploitation
• Monitoring for suspicious activity on ZMQ endpoints is recommended

Technical summary

The handle_zmq_recv coroutine in lmdeploy/pytorch/disagg/conn/engine_conn.py deserializes peer-to-peer cache-free requests with pickle.loads() without authentication, allowing remote code execution in deployments that enable disaggregated serving. This issue arises from the lack of authentication in the POST /distserve/p2p_initialize and /distserve/p2p_connect endpoints, which default to None unless the server is started with api_keys. A remote attacker can direct an engine to pull from a ZMQ endpoint under their control, leading to arbitrary code execution in the engine process. To address this vulnerability, it is crucial to assess exposure, apply patches or mitigations, and monitor for suspicious activity on

Defensive priority

High

Recommended defensive actions

• Review and apply patches from the vendor, specifically updating to version 0.16.0 or later
• Disable disaggregated serving if not required
• Implement authentication for API endpoints
• Monitor for suspicious activity on ZMQ endpoints
• Verify affected deployments and apply patches or mitigations
• Review compensating controls for exposed systems
• Track exceptions and retest remediated assets

Evidence notes

The CVE record and source references provide details on the vulnerability, including the affected code and potential impact. However, the corpus does not establish versions, exploitation, impact, or remediation for all affected deployments, requiring verification from the supplied official sources.

Sources and references

Methodology and review provenance

AI-assisted synthesis based on stored public vulnerability evidence. System validation, approval state, and publication status do not by themselves establish human review of this revision. PatchSiren helps prioritize defensive review and does not prove exposure or remediation on any system.