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-2026-103761 Overview
CVE-2026-103761 is a memory exhaustion vulnerability in the Mooncake transfer engine through version 0.3.13.post1. The flaw resides in the TransferMetadata::receivePeerNotify function, which processes notify frames on the handshake Remote Procedure Call (RPC) port. The function appends inbound notifications into an uncapped notifys vector without authentication or rate limiting. Unauthenticated remote attackers can repeatedly send notify frames up to 1 MB each, growing process memory until the operating system out-of-memory (OOM) killer terminates the engine. The weakness is classified as CWE-770: Allocation of Resources Without Limits or Throttling.
Critical Impact
Unauthenticated network attackers can crash the Mooncake transfer engine by exhausting host memory, disrupting any workload that depends on the service.
Affected Products
- Mooncake transfer engine versions up to and including 0.3.13.post1
- The kvcache-ai/Mooncake project distributions that expose the handshake RPC port
- Deployments embedding the vulnerable transfer_metadata.cpp logic
Discovery Timeline
- 2026-10-01 - CVE-2026-103761 published to the National Vulnerability Database (NVD)
- 2026-10-06 - Last updated in NVD database
Technical Details for CVE-2026-103761
Vulnerability Analysis
The Mooncake transfer engine coordinates key-value cache transfers between distributed inference workers. The TransferMetadata::receivePeerNotify handler listens on the handshake RPC port and accepts arbitrary notify frames from peers. Each frame is deserialized and pushed into an internal notifys vector that has no maximum size and no consumer-side back-pressure.
Because the handshake port does not require authentication, any network client that can reach the engine can submit frames. Each frame can be up to 1 MB in size, so a modest number of connections can quickly reserve gigabytes of resident memory. When the kernel OOM killer selects the engine process, inference traffic depending on the transfer engine halts.
For exact source context, see the referenced transfer_metadata.cpp lines 285-296 and the upstream GitHub Issue #4445.
Root Cause
The root cause is missing resource throttling in receivePeerNotify. The function appends into notifys without a size limit, per-peer quota, or rate limit. The handshake RPC port also lacks authentication, so no identity check gates the allocation. These two gaps combine to permit unbounded growth from untrusted sources.
Attack Vector
An attacker reaches the handshake RPC port over the network and repeatedly sends notify frames approaching the 1 MB per-frame ceiling. The engine deserializes each frame and appends it to the notifys vector. The attacker does not need credentials, user interaction, or prior access. Sustained traffic forces the host OOM killer to terminate the engine process, producing a denial-of-service condition. See the VulnCheck Advisory for the full technical write-up.
// No verified exploit code is published for CVE-2026-103761.
// The attack pattern is: open a TCP session to the handshake RPC port,
// send notify frames up to 1 MB in size in a loop, and observe RSS growth
// in the Mooncake transfer engine process until the OOM killer fires.
Detection Methods for CVE-2026-103761
Indicators of Compromise
- Rapid Resident Set Size (RSS) growth of the Mooncake transfer engine process without a matching increase in legitimate inference workload.
- Kernel log entries from oom-killer terminating the Mooncake transfer engine process.
- High inbound packet rate or large frame sizes directed at the handshake RPC port from unexpected source addresses.
Detection Strategies
- Monitor process-level memory metrics for the transfer engine and alert on sustained growth beyond baseline.
- Capture network flow telemetry for the handshake RPC port and flag traffic from hosts outside the inference cluster.
- Parse /var/log/kern.log or dmesg for Out of memory: Killed process events referencing the Mooncake binary.
Monitoring Recommendations
- Enable per-process memory and file descriptor telemetry on hosts running the Mooncake transfer engine.
- Forward kernel OOM events and container restart events to a central Security Information and Event Management (SIEM) platform.
- Baseline normal frame size distribution on the handshake RPC port and alert on sustained large-frame traffic.
How to Mitigate CVE-2026-103761
Immediate Actions Required
- Restrict network reachability to the handshake RPC port using host firewalls or Kubernetes NetworkPolicy, allowing only trusted peers.
- Place the Mooncake transfer engine behind an authenticated proxy or service mesh that enforces mutual TLS.
- Set container or systemd memory limits so a crash is contained and observable rather than destabilizing the host.
Patch Information
At the time of publication, the enriched advisory data does not list a fixed version. Monitor the upstream Mooncake repository and Issue #4445 for a patched release and apply it as soon as it becomes available. Review the referenced source lines in transfer_metadata.cpp to confirm whether the queue-bounding fix has been merged into your deployed version.
Workarounds
- Isolate the handshake RPC port on a dedicated management network segment unreachable from untrusted clients.
- Terminate connections through a reverse proxy that enforces maximum message size and per-source connection rate limits.
- Apply cgroup memory.max limits to the engine process to bound total consumption and trigger controlled restarts.
# Example: restrict the handshake RPC port to a trusted CIDR using iptables
# Replace 10.0.0.0/24 with the trusted peer subnet and 8080 with the actual port
iptables -A INPUT -p tcp --dport 8080 -s 10.0.0.0/24 -j ACCEPT
iptables -A INPUT -p tcp --dport 8080 -j DROP
# Example: cap memory for the engine using systemd
# /etc/systemd/system/mooncake-transfer-engine.service.d/override.conf
# [Service]
# MemoryMax=4G
# Restart=on-failure
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

