Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-106039 Overview
CVE-2026-106039 is a missing authorization vulnerability [CWE-862] in Mooncake Store master versions through 0.3.13.post1. The flaw exists in the coro_rpc port, which exposes replication task management functions without authentication checks. Unauthenticated network attackers can invoke CreateCopyTask, CreateMoveTask, FetchTasks, and MarkTaskToComplete using victim client UUIDs disclosed by QueryTask. This allows adversaries to create, steal, and falsely mark replication tasks as complete, hijacking task queues and recording replication events that never occurred.
Critical Impact
Attackers with network access to the coro_rpc port can tamper with distributed KV cache replication state in Mooncake Store deployments used for large language model (LLM) serving infrastructure.
Affected Products
- Mooncake Store master branch through version 0.3.13.post1
- Deployments exposing the coro_rpc port to untrusted networks
- LLM serving environments using Mooncake for distributed KV cache replication
Discovery Timeline
- 2026-10-06 - CVE-2026-106039 published to NVD
- 2026-10-06 - Last updated in NVD database
Technical Details for CVE-2026-106039
Vulnerability Analysis
Mooncake Store is a distributed KV cache store used in LLM inference infrastructure. The master service manages replication tasks across cache nodes using remote procedure calls over the coro_rpc transport. The vulnerability stems from the absence of caller identity verification on four RPC endpoints that control the replication task lifecycle.
An attacker reachable on the network can first call QueryTask to enumerate active client UUIDs. With a valid UUID in hand, the attacker can then issue CreateCopyTask or CreateMoveTask RPCs that masquerade as the victim client. The same identity-spoofing path enables calls to FetchTasks, draining the victim's queue, and MarkTaskToComplete, which signals success without any replication actually occurring.
The practical consequence is corruption of replication state. Downstream consumers believe cache objects have been copied or moved when they have not, leading to silent data integrity issues within the distributed cache.
Root Cause
The affected RPC handlers in master_service.cpp and task_manager.cpp accept client UUIDs as untrusted parameters without binding them to an authenticated session. Any caller that supplies a known UUID is treated as the legitimate owner of the associated tasks. This is a classic Missing Authorization condition classified under [CWE-862].
Attack Vector
Exploitation requires only network reachability to the coro_rpc listener. No authentication, user interaction, or elevated privileges are needed. The attack proceeds as follows:
- The attacker connects to the exposed coro_rpc port on the Mooncake master.
- QueryTask is invoked to enumerate active client UUIDs.
- Using a stolen UUID, the attacker calls CreateCopyTask or CreateMoveTask to inject rogue entries.
- FetchTasks is called to steal pending work from the victim client.
- MarkTaskToComplete is invoked to falsely signal completion of replication operations.
For technical details, see the VulnCheck Advisory for Mooncake Store and the referenced code in the GitHub Mooncake Master Service Code.
Detect Identity Attacks the Moment Authentication Fails
Singularity™ Identity surfaces credential misuse, lateral movement, and privilege escalation in real time, so you can contain the breach before it spreads.
Stop Identity AttacksDetection Methods for CVE-2026-106039
Indicators of Compromise
- Unexpected CreateCopyTask or CreateMoveTask RPC calls originating from IP addresses outside the known cache node fleet
- Repeated QueryTask invocations followed by FetchTasks calls referencing enumerated UUIDs
- Replication tasks reaching the completed state without corresponding data transfer metrics
- Divergence between master-side task completion logs and worker-side replication activity
Detection Strategies
- Monitor coro_rpc access logs for unauthenticated callers reaching replication task endpoints
- Correlate task lifecycle events with network flow records to identify RPC calls from unexpected sources
- Baseline normal client UUID activity patterns and alert on anomalous enumeration behavior
Monitoring Recommendations
- Capture network telemetry on the coro_rpc port and forward to a centralized analytics platform
- Instrument the Mooncake master to emit structured audit events for every replication RPC
- Alert on replication tasks marked complete with zero bytes transferred or with mismatched source and destination counts
How to Mitigate CVE-2026-106039
Immediate Actions Required
- Restrict network access to the coro_rpc port using host-based firewalls or network segmentation so only trusted cache nodes can connect
- Audit recent replication task logs for evidence of UUID-based hijacking or falsely completed tasks
- Isolate Mooncake Store deployments from untrusted or shared network segments until a patched release is applied
Patch Information
At the time of publication, Mooncake Store versions through 0.3.13.post1 remain affected. Track GitHub Mooncake Issue #4475 and the GitHub Mooncake Repository for the fix release. Upgrade to a patched version as soon as it is available.
Workarounds
- Place the coro_rpc listener behind a mutually authenticated transport such as mTLS using an upstream proxy
- Bind the coro_rpc port to a private interface or management VLAN rather than a public address
- Deploy network ACLs restricting inbound connections to known Mooncake worker node IP ranges
- Monitor and rate-limit QueryTask calls to slow UUID enumeration attempts
# Example iptables rule restricting coro_rpc port to trusted cache nodes
# Replace <CORO_RPC_PORT> and trusted subnet with environment-specific values
iptables -A INPUT -p tcp --dport <CORO_RPC_PORT> -s 10.0.0.0/24 -j ACCEPT
iptables -A INPUT -p tcp --dport <CORO_RPC_PORT> -j DROP
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

