CVE-2026-106037: Mooncake Distributed KVCache Store Unauthenticated REST API Exposure and Cache Poisoning
HERMES THREAT SCORE & AGENTIC INFRASTRUCTURE EVALUATION
Target:Mooncake Distributed KVCache Store โ Store REST Service (mooncake_store_service.py) Both CVSS v4.0 (9.3 Critical) and the Hermes Threat Score (94 CRITICAL, with an Agentic Security Score HASS of 95) highlight a fundamental boundary failure in distributed LLM serving architecture. Mooncake powers high-throughput inference for large models (vLLM and Kimi disaggregated prefill-decode clusters) by serving as a distributed Key-Value (KV) cache fabric. In versions up to 0.3.13.post1, the Store HTTP service unconditionally binds to 0.0.0.0 on port 8080 without requiring any authentication tokens, mutual TLS, or network access controls. Remote adversaries on routable networks can invoke /api/get to reconstruct raw prompt contexts and chat histories, call /api/put to inject adversarial KV tensor slices directly into running inference jobs (MITRE ATLAS AML.T0043), or purge memory segments via /api/remove_all to trigger catastrophic cluster-wide denial of service.
CVE-2026-106037: Mooncake Distributed KVCache Store Unauthenticated REST API Exposure and Cache PoisoningVULNERABILITY
Software platform affected by security vulnerabilities and agentic attack patterns.
๐ Why is this related? (Evidence & Provenance)
“Confirmed security vulnerability in Mooncake Distributed KVCache Store documented in Hermes dossier.”
- [vulnerability_report]
- [government_confirmation]CISA verified active exploitation in the wild and mandated federal remediation deadline in KEV entry. — Source: Cybersecurity & Infrastructure Security Agency (CISA): CISA Adds CVE-2026-59822 to Known Exploited Vulnerabilities Catalog (Reliability: VERY_HIGH)
1. Technical Context & Affected Architecture Matrix
Section titled โ1. Technical Context & Affected Architecture MatrixโModern multi-node LLM serving separates compute-intensive prompt prefilling from token generation (decode). Mooncake acts as the decentralized memory pooling orchestrator, coordinating RDMA, NVMe, and DRAM cache segments.
| Parameter | Technical Specification | Operational Details |
|---|---|---|
| CVE Identifier | CVE-2026-106037 | International security tracking identifier |
| Software Platform | Mooncake KVCache Store | Open-source distributed KVCache transfer engine for LLMs |
| Vulnerable Component | mooncake-wheel/mooncake/mooncake_store_service.py | Store REST Service (start_http_service()) |
| Vulnerability Class | CWE-306 / CWE-284 / CWE-200 | Missing Authentication for Critical Function / Information Disclosure |
| Attack Vector | Network (Port 8080 TCP) | Unauthenticated HTTP REST requests (GET, PUT, POST, DELETE) |
| Remediation Release | Mooncake v0.3.14 | Mandatory bearer token validation and loopback binding defaults |
[ Unauthenticated Remote Attacker ] | Direct HTTP requests to :8080 TCP No bearer tokens, API keys, or mTLS | v +------------------------------------------+ | Mooncake Store REST API (:8080 TCP) | | aiohttp Web Server (0.0.0.0 binding) | +------------------------------------------+ / | \ GET /api/get/{key} / | \ PUT /api/put v v v +-------------------+ +-------------------+ +-------------------+ | Exfiltrate Prompt | | DELETE /api/ | | Inject Poisoned | | Cache Tensors | | remove_all | | KV Tensor Slices | | & System Contexts | | Cluster DoS | | (AML.T0043) | +-------------------+ +-------------------+ +-------------------+ | | v v +------------------------------------------+ | Disaggregated LLM Inference Pool | | Decode Worker loads poisoned tensors | | Agentic Reasoning Compromised! | +------------------------------------------+2. Root Cause Analysis & Vulnerable Code Anatomy
Section titled โ2. Root Cause Analysis & Vulnerable Code AnatomyโThe root cause resides in mooncake-wheel/mooncake/mooncake_store_service.py. When initializing the REST management server, start_http_service() creates an aiohttp application exposing high-privilege operations with zero middleware verification:
# Vulnerable implementation in mooncake_store_service.py (Mooncake <= 0.3.13.post1)async def start_http_service(self, port: int = 8080): app = web.Application(client_max_size=1024 * 1024 * 100) # 100MB limit app.add_routes( [ web.post("/api/reconfigure", _timed_handler("RECONFIGURE", self.handle_reconfigure)), web.post("/api/mount_shm", _timed_handler("MOUNT_SHM", self.handle_mount_shm)), web.post("/api/unmount_shm", _timed_handler("UNMOUNT_SHM", self.handle_unmount_shm)), web.post("/api/mount", _timed_handler("MOUNT", self.handle_mount)), web.post("/api/unmount", _timed_handler("UNMOUNT", self.handle_unmount)), web.post("/api/unmount_local_disk", _timed_handler("UNMOUNT_LOCAL_DISK", self.handle_unmount_local_disk)), web.put("/api/put", _timed_handler("PUT", self.handle_put)), web.get("/api/get/{key}", _timed_handler("GET", self.handle_get)), web.get("/api/exist/{key}", _timed_handler("EXIST", self.handle_exist)), web.delete("/api/remove/{key}", _timed_handler("REMOVE", self.handle_remove)), web.delete("/api/remove_all", _timed_handler("REMOVE_ALL", self.handle_remove_all)), ] ) runner = web.AppRunner(app) await runner.setup() # FLAW: Binds unconditionally to 0.0.0.0 without authentication middleware site = web.TCPSite(runner, "0.0.0.0", port) await site.start() logging.info(f"REST API started on port {port}") return TrueExploit Mechanics:
Section titled โExploit Mechanics:โ- Zero Route Authorization: None of the route handlers verify caller identity or session tokens.
- Wildcard Interface Binding: The service binds to
0.0.0.0, exposing internal GPU cluster communications to any adjacent network or internet gateway. - Arbitrary KV Access: KV caches represent linearized attention tensors corresponding to tokenized input prompts. Dumping the keys allows rebuilding the cleartext prompt sequence.
3. Attack Vectors & Real-World Exploitation Scenarios
Section titled โ3. Attack Vectors & Real-World Exploitation ScenariosโVector A: Prompt Cache Harvesting & Model Theft
Section titled โVector A: Prompt Cache Harvesting & Model TheftโAdversaries query the key enumeration or existing known keys to download serialized tensor buffers:
# Harvest active session KV cache tensorcurl -s -X GET "http://target-inference-node.internal:8080/api/get/session_ctx_94821_layer_0" \ -o stolen_tensor_l0.binUsing a reverse tokenizer mapping, the adversary converts cached embeddings back into sensitive inputs (corporate contracts, proprietary source code, or internal database queries passed to autonomous agents).
Vector B: Adversarial Tensor Poisoning (MITRE ATLAS AML.T0043)
Section titled โVector B: Adversarial Tensor Poisoning (MITRE ATLAS AML.T0043)โAdversaries overwrite an active sessionโs KV cache with crafted tensor slices that steer agent decision-making:
# Inject malicious KV tensor into victim agent sessioncurl -s -X PUT "http://target-inference-node.internal:8080/api/put" \ -H "Content-Type: application/octet-stream" \ -H "X-Key: session_ctx_94821_layer_0" \ --data-binary @crafted_backdoor_kv.binWhen the decode node resumes token generation, the attention mechanism incorporates the forged context, forcing the agent to invoke privileged tools or exfiltrate credentials without the prompt text ever appearing in front-end audit logs.
4. Forensic Triage & Behavioral Signatures
Section titled โ4. Forensic Triage & Behavioral Signaturesโtitle: Mooncake Distributed KVCache Store Unauthenticated REST Invocationid: c7143e59-8641-4562-b9b2-3c89b71e12aastatus: experimentaldescription: Detects unauthenticated or unauthorized HTTP requests targeting Mooncake distributed KVCache store endpoints.references: - https://github.com/kvcache-ai/Mooncake/issues/4476 - https://vulncheck.com/advisories/mooncake-through-0.3.13-post1-missing-authentication-in-store-rest-serviceauthor: Hermes Codex CTI Teamdate: 2026-10-07logsource: category: webserver product: aiohttpdetection: selection_endpoints: cs-method: - 'GET' - 'PUT' - 'POST' - 'DELETE' cs-uri-stem|startswith: - '/api/get/' - '/api/put' - '/api/remove_all' - '/api/mount' - '/api/reconfigure' filter_internal: c-ip: - '127.0.0.1' - '::1' condition: selection_endpoints and not filter_internalfields: - c-ip - cs-method - cs-uri-stem - sc-statusfalsepositives: - Legitimate cluster health monitors or authorized non-loopback management if configuredlevel: criticaltags: - attack.initial_access - attack.exfiltration - atlas.aml.t0043// Azure Sentinel / Defender for Cloud: Detect Ingress to Mooncake KVCache EndpointsCommonSecurityLog| where DestinationPort == 8080| where RequestMethod in ("GET", "PUT", "POST", "DELETE")| where RequestURL has_any ("/api/get", "/api/put", "/api/remove_all", "/api/mount", "/api/reconfigure")| project TimeGenerated, SourceIP, DestinationIP, DestinationPort, RequestMethod, RequestURL, DeviceAction| extend SuspiciousAction = case( RequestURL has "/api/remove_all", "Cache Purge / Denial of Service", RequestURL has "/api/put", "Adversarial Tensor Injection", RequestURL has "/api/get", "KV Cache Exfiltration", "Store Reconfiguration")| order by TimeGenerated desc#!/usr/bin/env bash# Audit Mooncake Store listening interfaces and active network connectionsecho "[*] Auditing Mooncake listening ports..."ss -tlpn | grep ":8080"
echo "[*] Checking for external established connections..."ss -tan state established '( dport = :8080 or sport = :8080 )'
echo "[*] Verifying Mooncake process arguments..."ps aux | grep "[m]ooncake_store_service"5. Defense-in-Depth & Remediation Strategy
Section titled โ5. Defense-in-Depth & Remediation Strategyโ- Upgrade Mooncake: Update the
mooncakePython package and container runtimes to version 0.3.14 or later. - Restrict Network Binding: Configure Mooncake service parameters to listen strictly on
127.0.0.1or the designated private RDMA cluster interface:Terminal window export MOONCAKE_STORE_HOST="127.0.0.1"export MOONCAKE_STORE_PORT="8080" - Enforce API Token Authentication: Enable bearer token authentication in Mooncake configuration files, requiring
Authorization: Bearer <CLUSTER_SECRET>on all REST routes. - Network Segmentation: Implement Kubernetes
NetworkPolicyrules isolating port 8080 TCP strictly between prefill pods and decode pods, dropping all egress/ingress from general pod networks.