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-2026-103765 Overview
CVE-2026-103765 is a missing authentication vulnerability [CWE-306] in Mooncake, an open-source key-value (KV) cache and transfer engine used in large language model (LLM) inference pipelines. Versions through 0.3.13.post1 expose the /metadata handler on the HTTP metadata server without any authentication controls. Unauthenticated network attackers can read, overwrite, and delete transfer engine metadata keys. By poisoning segment descriptors such as tcp_data_port or recreating rpc_meta entries, an attacker redirects KV cache transfers to attacker-controlled listeners or exhausts server memory.
Critical Impact
Unauthenticated attackers can intercept KV cache data transfers between inference nodes, resulting in confidentiality loss and denial of service across Mooncake-backed LLM infrastructure.
Affected Products
- Mooncake versions through 0.3.13.post1
- Mooncake HTTP metadata server component (http_metadata_server.py)
- LLM inference deployments that rely on Mooncake transfer engine metadata
Discovery Timeline
- 2026-10-02 - CVE-2026-103765 published to NVD
- 2026-10-02 - Last updated in NVD database
Technical Details for CVE-2026-103765
Vulnerability Analysis
Mooncake provides a disaggregated KV cache transfer engine that coordinates inference nodes through a metadata service. The HTTP metadata server exposes a /metadata endpoint that stores segment descriptors and RPC coordination entries used to route KV cache traffic between workers.
The /metadata handler accepts GET, PUT, and DELETE operations without validating the identity of the caller. Any network-reachable client can enumerate keys, overwrite entries such as tcp_data_port, or recreate rpc_meta records. Downstream workers then connect to attacker-specified endpoints when transferring KV cache contents, exposing inference data and model-side state to interception.
The same primitive supports a denial-of-service path. An attacker can insert large or unbounded metadata entries to exhaust server memory, degrading or halting the inference pipeline. See the VulnCheck Security Advisory for additional detail.
Root Cause
The root cause is the absence of authentication and authorization checks on the /metadata handler. The handler, implemented in http_metadata_server.py, trusts any client that can reach the HTTP port. Mooncake assumes the metadata service runs inside a trusted network boundary, but the service does not enforce that assumption.
Attack Vector
Exploitation requires only network access to the Mooncake metadata server HTTP port. An attacker issues a PUT request to /metadata that overwrites a target segment key, substituting attacker-controlled values for fields such as tcp_data_port. Worker nodes then direct KV cache transfers to the attacker's listener. Alternatively, repeated PUT requests with large payloads exhaust server memory. See the Mooncake project repository and GitHub Issue #4444 for upstream discussion.
No verified public exploit code is available at this time. The vulnerability manifests in the metadata handler described above; refer to the linked advisory for technical details.
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-2026-103765
Indicators of Compromise
- Unexpected PUT or DELETE requests to the /metadata endpoint from clients outside the inference cluster.
- Metadata entries where tcp_data_port or rpc_meta fields reference IP addresses or hostnames outside the known worker pool.
- Sudden growth in metadata server memory consumption without a corresponding increase in legitimate workload.
- KV cache transfer connections established to endpoints not listed in the deployment inventory.
Detection Strategies
- Enable HTTP access logging on the Mooncake metadata server and alert on write operations from source addresses outside an allowlist.
- Baseline the set of metadata keys and segment descriptors, then alert on unexpected key creation or value changes.
- Correlate inference worker outbound connections with the current metadata view to identify redirection to unknown hosts.
Monitoring Recommendations
- Monitor process memory on the metadata server and alert on sustained growth beyond the normal operating envelope.
- Instrument the transfer engine to log the resolved tcp_data_port for every KV transfer and flag non-cluster destinations.
- Capture network flow data between the metadata server and workers for post-incident reconstruction.
How to Mitigate CVE-2026-103765
Immediate Actions Required
- Restrict network access to the Mooncake metadata server HTTP port to inference cluster members using firewall rules, security groups, or Kubernetes NetworkPolicy objects.
- Remove the metadata server from any interface reachable by untrusted networks, including shared internal subnets.
- Audit current metadata entries for unexpected tcp_data_port or rpc_meta values and remove any that do not match known workers.
- Rotate any model or session data that may have transited through compromised transfer paths.
Patch Information
No fixed version is identified in the published advisory at the time of writing. Mooncake users should track the upstream Mooncake repository and GitHub Issue #4444 for a release that introduces authentication on the /metadata handler. Until a patched version is available, apply the network and workaround controls below.
Workarounds
- Bind the metadata server to a loopback or management-only interface and front it with an authenticating reverse proxy that enforces mutual TLS or token authentication.
- Place the metadata server behind a service mesh that applies identity-based access policies to the /metadata route.
- Deploy the Mooncake control plane on a dedicated VLAN or private subnet isolated from workload and user traffic.
- Apply rate limits and request-size limits at the proxy layer to blunt memory exhaustion attempts.
# Example iptables rule limiting /metadata port (default 8080) to a known worker CIDR
iptables -A INPUT -p tcp --dport 8080 -s 10.0.10.0/24 -j ACCEPT
iptables -A INPUT -p tcp --dport 8080 -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

