Summary
CVE-2026-103765 is a critical missing-authentication vulnerability in Mooncake's HTTP metadata server. An unauthenticated attacker may read, overwrite, or delete transfer-engine metadata, potentially redirecting KV cache transfers to attacker-controlled listeners or exhausting server memory.
Affected context
The issue occurs in Mooncake deployments through 0.3.13.post1 when the HTTP metadata server exposes the /metadata handler without an effective authentication and authorization boundary. It is especially risky when the handler is reachable from untrusted networks or when metadata controls transfer-engine routing and resource-management behavior.
Exploit preconditions
- Attack surface: The HTTP metadata server and its /metadata handler - Required access: Network reachability to the handler; valid credentials
may not be required
- Trigger point: Requests that read, overwrite, or delete transfer-engine
metadata keys
- Why exploitation is possible: The handler accepts sensitive metadata
operations without adequately verifying the caller's identity and
permissions
Why this happens
The underlying bug is a missing authentication control on an administrative or control-plane HTTP endpoint. The application permits security-sensitive metadata operations before establishing that the requester is an authorized principal, which maps to CWE-306, Missing Authentication for Critical Function. Because metadata can influence transfer destinations and resource state, unauthenticated modification can affect both confidentiality and availability.
Potential impact
- Attackers may read sensitive transfer-engine metadata
- Attackers could overwrite segment descriptors and redirect KV cache transfers
- Attackers may recreate RPC metadata entries that influence transfer routing
- The service could experience memory exhaustion and denial of service
- Transferred data could be exposed to attacker-controlled listeners
General mitigation guidance
Mitigation generally focuses on placing the metadata interface behind strong, mandatory authentication and fine-grained authorization. Sensitive read, write, and delete operations should use separate least-privilege permissions, while metadata values should be validated against strict schemas and trusted endpoint policies. Exposure can be reduced through network isolation, authenticated service-to-service communication, and secure-by-default binding and access-control settings. Runtime identities should have only the permissions required for metadata management, and untrusted clients should not be able to influence transfer destinations.
How teams detect issues like this
Detection can include static and configuration analysis for HTTP handlers that expose state-changing operations without authentication middleware, along with code-review checks for missing authorization around metadata reads and mutations. Runtime monitoring should alert on unauthenticated requests, unexpected metadata changes, unusual endpoint exposure, and transfer destinations that fall outside approved policies. Access-control policies and automated review tools can help enforce authentication requirements for control-plane interfaces.
