Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata – not a confirmed exploit chain
-
AccessReach Mooncake TCP transport data port
-
DeliverySend crafted SessionHeader with arbitrary addr/size
-
ExploitTrigger untrusted pointer dereference in readHeader
-
ExecutionRead or write arbitrary process memory
-
ImpactExfiltrate KV cache/secrets or corrupt memory for code execution
Vulnerability Assessment
| Weakness |
CWE
ASVS 5.0
V15 - Secure Coding and Architecture Untrusted pointer prevention
|
| Exploitation Conditions | Exploitation requires network reachability to the Mooncake transfer engine's TCP transport data port (the socket handling SessionHeader messages in ServerSession::readHeader); the attacker then sends a SessionHeader with attacker-chosen addr and size fields using a READ or WRITE opcode. No authentication, credentials, or user interaction are needed (PR:N/UI:N), and default deployments of the TCP transport expose this port. The practical limiting factor is network exposure: Mooncake data ports are typically bound within a trusted LLM-serving cluster/back-end network, so risk is highest where that port is reachable by untrusted hosts (flat networks, misconfigured segmentation, or internet exposure) and lower where the data plane is isolated. Only versions before 0.3.13 are affected. |
| Impact | This is a genuine high-priority vulnerability, not a high-CVSS-but-low-real-risk case. The CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:N/UI:N) describes a network-reachable, low-complexity flaw requiring no authentication, no special attack requirements, and no user interaction, with high confidentiality, integrity, and availability impact (VC:H/VI:H/VA:H) - the maximum exploitability profile short of a scope change. CWE-822 (untrusted pointer dereference) combined with the description's claim of attacker-controlled addr/size in READ or WRITE opcodes means an unauthenticated attacker gets a primitive for arbitrary process-memory read AND write, which is directly weaponizable toward both data disclosure (KV cache, prompts, secrets) and code execution. The flaw is corroborated by independent source-code evidence: a vendor patch commit (a2933849) and the vulnerable code path in tcp_transport.cpp, and it was reported by a reputable research vendor (VulnCheck) with its own advisory. Key caveats: no EPSS score was provided, so exploitation probability cannot be quantified; there is no CISA KEV listing and no public exploit code referenced, so there is no evidence of active exploitation - the 'RCE' tag reflects potential impact, not observed in-the-wild attacks. The CVSS network vector is fully consistent with the description (TCP transport data port), so there is no signal conflict; the primary unknown is real-world exposure of the data port, which is typically on an internal cluster network rather than the public internet. |
| Exploit ScenarioDerived | An attacker with network reachability to a Mooncake transfer engine's TCP transport data port connects and sends a crafted SessionHeader carrying an arbitrary memory address and size with a READ opcode, causing ServerSession::readHeader to dereference the untrusted pointer and return the contents of that memory region - disclosing cached KV data, prompts, or secrets. Using a WRITE opcode with a controlled address, the same primitive corrupts process memory to pivot toward remote code execution. No public exploit code is referenced in the available data, but the vendor patch and VulnCheck advisory confirm the bug's existence and exploitability. |
Get an e-mail when this CVE is exploited, patched or re-scored. Free account.
Alert meAnalysisAIv1 – Oct 01, 2026
Unauthenticated attackers with network reachability to the Mooncake transfer engine's TCP transport data port on releases before 0.3.13 can achieve arbitrary process memory read and write by sending a crafted SessionHeader whose addr and size fields are attacker-controlled under a READ or WRITE opcode. Successful exploitation discloses in-memory KV cache contents, prompts and other secrets, and corrupts memory in the transfer engine process, giving a credible route toward further compromise of the served model; the 9.3 score reflects a network-reachable, no-credentials, no-interaction attack with high confidentiality, integrity and availability impact confined to the vulnerable component. No public exploit code has been identified at time of analysis, and the practical limiting factor is exposure of the data port, which is normally bound inside a trusted LLM-serving cluster network - flat, misconfigured or internet-facing deployments carry the highest risk. Vendor-released patches are available in v0.3.13.
RemediationAI
Upgrade the Mooncake transfer engine to v0.3.13 or later (vendor-released patch; upstream fix commit a2933849417259e562fc9cbad63c618d464dfb2d) and confirm every node in the serving cluster - prefill, decode and any standalone transfer-engine processes - is rebuilt or restarted against the patched library, since the engine is typically linked into the model server process and an un-restarted worker remains exploitable. If immediate upgrade is not possible, the only effective compensating control is network restriction, because the TCP transport performs no authentication: allow the data port only from known peer worker IPs using host firewalls, security groups or cluster network policy, and place the data plane on a dedicated, isolated segment so untrusted or flat-network hosts cannot reach it. Where RDMA transports are available and used, disabling the TCP transport avoids the vulnerable parsing path entirely, at the cost of losing the TCP fallback for hosts without RDMA and potentially breaking connectivity for mixed clusters. Exposing the data port to the internet or to shared multi-tenant networks should be treated as unacceptable pre-patch. Detection is difficult from logs alone, so log all accepted connections to the data port during the exposure window and treat any unexpected source IP as a potential memory disclosure event; review the vendor advisory at https://www.vulncheck.com/advisories/mooncake-transfer-engine-before-0.3.13-unauthenticated-arbitrary-memory-read-write-via-tcp-transport and the upstream issue at https://github.com/kvcache-ai/Mooncake/issues/4441 for further detail.
Affected VersionsENISA
Affected Packages
Source: OSV (Open Source Vulnerabilities)
Affects your stack
Affects 1 package in euvd
Checking this CVE against your own stacks and servers is part of the Basic plan and up.
Technical ContextAI · show details
Mooncake (kvcache-ai) is a KVCache-centric data transfer engine used in disaggregated LLM serving, where prefill and decode workers exchange KV cache blocks and prompts over a data plane that can use RDMA or, as a fallback, a raw TCP transport. The vulnerable code path is ServerSession::readHeader in mooncake-transfer-engine/src/transport/tcp_transport/tcp_transport.cpp (referenced line 209), which deserializes a SessionHeader from the TCP data socket and then uses the header's addr and size values as the destination/source pointer and transfer length. Because those fields are taken directly from the wire without validation against the registered/allocated memory regions, the flaw is a classic CWE-822 untrusted pointer dereference: a READ opcode returns process memory at the attacker's chosen address and a WRITE opcode writes attacker-supplied bytes to that address. The protocol performs no authentication or authorization, so any host that can open a connection to the data port and speak the framing is a valid client. The fix landed in the upstream commit a2933849417259e562fc9cbad63c618d464dfb2d and is published in release v0.3.13, which reworks the TCP transport's session/connection handling (bounded work queues, connection lanes and session lifecycle management) and addresses the unsafe header handling. Note that the independently assessed impact vector keeps scope unchanged (S:U), so the confirmed impact is full read/write within the transfer engine process; code execution or escape to other components is plausible but not established by the available data.
Know when Mooncake gets its next vulnerability
- Watch products and get an email when something new is published
- Check this CVE against your own stack and servers
- Exploit and PoC references, threat intelligence from 22 sources
No credit card · 7-day full trial
Severity by source
CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N
Unauthenticated network access to the TCP data port with no user interaction (AV:N/AC:L/PR:N/UI:N); arbitrary memory read/write yields full C/I/A impact within the unchanged process scope.
Primary rating from Vendor (VulnCheck).
CVSS VectorVendor: VulnCheck
DescriptionCVE.org
Mooncake transfer engine before 0.3.13 contains an untrusted pointer dereference in ServerSession::readHeader that allows unauthenticated attackers to read and write arbitrary process memory via the TCP transport data port. Attackers can send a crafted SessionHeader with arbitrary addr and size values using READ or WRITE opcodes to disclose KV cache contents, prompts and secrets or corrupt memory toward code execution.
Executive ViewAI
The Mooncake transfer engine, in releases prior to 0.3.13, exposes a memory read and write flaw that allows anyone who can reach its TCP data port to extract in-memory KV cache contents, prompts and other secrets, and to corrupt the serving process. Because that port is normally bound inside a trusted LLM-serving cluster network, the organizations most at risk are those running flat, misconfigured or internet-facing inference deployments where that boundary has eroded. Leadership should treat this as an exposure and containment problem: the blast radius is confined to affected inference infrastructure, but the data at risk is precisely the sensitive model context and user prompts the business is accountable for protecting.
Apply vendor patch.
Threat Intelligence Overview
How this decision was reached
The engine reads each signal below and applies a fixed list of rules from top to bottom. The first rule that matches sets the recommended action – like a statement, every line is shown.
What every action level means
SSVC AssessmentENISA
Priority Score
Attack Technique
Remote Code Execution represents the critical moment when an attacker successfully runs arbitrary code on a target system without physical access.
Unauthenticated access to Mooncake's HTTP metadata server (all versions through 0.3.13.post1) lets any network-reachable
Unauthenticated memory exhaustion in the Mooncake transfer engine (all versions through 0.3.13.post1) allows any network
Denial of service in Mooncake Transfer Engine through 0.3.13.post1 allows unauthenticated remote attackers to permanentl
Uncontrolled resource allocation in the MasterService::GetReplicaListByRegex handler of kvcache-ai Mooncake (all version
Improper access control in kvcache-ai Mooncake up to versions 0.3.12, 0.3.13.post1, and 0.3.14-rc1 allows remote attacke
EUVD-2026-91136
GHSA-mp8g-r7h2-5x36
NVD
CVE.org
EPSS
Vulnrichment
VulDB
Exploit-DB