WalledAI enterprise logo
Back to blog
AI SecurityOctober 2, 20267 min read

Mooncake Flaw Lets Attackers Read AI Memory Without Login

A critical Mooncake AI vulnerability allows unauthenticated remote attackers to read or write arbitrary memory, exposing LLM prompts, cached data and secrets. Security teams should immediately inventory AI serving infrastructure and restrict network exposure.

WA

Walled AI Research Team

On this page›

Mooncake, a high-performance data transfer engine used to serve large language models, contains a critical unauthenticated arbitrary memory read/write flaw. An attacker who can reach the service on its TCP data port can supply arbitrary memory addresses and sizes to read prompts, cached KV-cache data, secrets or corrupt process memory without credentials. The issue, tracked as CVE-2026-103764, is one of at least seven related vulnerabilities disclosed in early October 2026 and demands immediate inventory and network segmentation for AI serving infrastructure.

What happened and how does the flaw work?

The core issue lies in Mooncake's TCP transfer engine. In versions before 0.3.13, the data plane implemented a raw memory-access protocol that accepted a network-supplied addr value directly as a process pointer, while size was not bounded against registered buffers. According to the maintainer's disclosure on GitHub, the TCP listener selects a port in the documented 15000-17000 range and binds on all interfaces. A remote attacker can send a crafted SessionHeader with arbitrary addr and size values using READ or WRITE operations, resulting in arbitrary process memory access. The researcher confirmed this behavior at runtime on version 0.3.11.post1, including reads and writes to both registered and unregistered memory (GitHub Issue #4441).

Mooncake v0.3.13 was released on August 26, 2026, and its release notes include a transfer engine change titled "Harden socket handshake frame handling." The vulnerability report identifies the relevant fix commit and states that 0.3.13 added validateTcpAddress, which rejects addresses outside registered buffers. However, the same report explicitly notes that the TCP port still lacks peer authentication, meaning the channel remains unauthenticated even after the address validation patch (v0.3.13 release notes; GitHub Issue #4441).

Why does this matter for CISOs and GRC teams?

The CNA description for CVE-2026-103764 states that consequences include disclosure of KV-cache contents, prompts and secrets, memory corruption and possible progression toward code execution. The vulnerability carries a CVSS 3.1 score of 9.8 Critical and a CVSS 4.0 score of 9.3 Critical, with a network attack vector, low complexity, no privileges required and no user interaction (VulnCheck advisory; Rapid7 vulnerability database).

For enterprise security and governance leaders, this is not a conventional application bug. Mooncake sits in the data path that moves the tensors, prompts and cached state needed to serve large language models. An unauthenticated memory read primitive can expose live prompts that contain personally identifiable information, regulated healthcare or financial data, or corporate trade secrets. Memory-resident secrets such as cloud provider API keys or database credentials are also within reach. Because the flaw permits writes as well as reads, an attacker could corrupt memory to achieve code execution inside the AI serving environment, potentially pivoting to adjacent systems or poisoning model outputs.

As of October 2, 2026, there is no CISA Known Exploited Vulnerabilities entry and no EPSS score available for this flaw. Rapid7 and SANS Internet Storm Center both note this absence, but it is important to treat it as a reflection of database timing rather than evidence that exploitation has not occurred (Rapid7; SANS Internet Storm Center). Security teams should assume active scanning and proof-of-concept release are likely.

The exposure of live prompts and secrets also raises AI data loss prevention concerns. Organizations that have invested in prompt-level controls must now verify that the infrastructure transporting those prompts is not undermining them with unauthenticated network interfaces.

What are the risks beyond the headline flaw?

The October 2026 disclosure is a cluster, not a single bug. OpenCVE currently lists seven Mooncake CVEs, and at least three distinct security problems deserve separate treatment.

First, CVE-2026-103765 concerns the HTTP metadata server rather than the TCP transport. The /metadata endpoint reportedly permits GET, PUT and DELETE without authentication. A GET request can expose segment descriptors containing absolute buffer addresses, TCP data ports and RDMA rkeys. More critically, an attacker can overwrite descriptors to redirect live KV-cache transfers to an attacker-controlled listener while the sender still reports success. The issue was reportedly still present in version 0.3.13 and the main branch when the report was opened, with the researcher stating that no fix was available at that time (GitHub Issue #4444).

Second, CVE-2026-103761 is a denial-of-service flaw in the handshake daemon. The daemon appends every received notify message to an in-memory vector with no cap, expiry or per-peer quota. Because a notify message can be up to roughly 1 MB under the default MC_HANDSHAKE_MAX_LENGTH, memory consumption can grow at roughly one byte of resident set size per byte sent. The vLLM Mooncake connector reportedly does not call the function that drains the queue. This affects versions up to and including 0.3.13.post1 and carries a CVSS 4.0 score of 8.7 (GitHub Issue #4445; OpenCVE CVE-2026-103761).

The practical implication is that simply finding version 0.3.13 in an environment does not mean the organization is clear of the disclosure cluster. Security teams must map each exposed interface to the specific CVEs that affect it and track fixes for the metadata server and handshake daemon separately from the arbitrary-memory primitive.

Is version 0.3.13 enough to stop the attacks?

Many vulnerability databases list Mooncake versions before 0.3.13 as affected by CVE-2026-103764, and the maintainer's release notes for 0.3.13 mention hardening the socket handshake frame handling. The technical report confirms that this release added validateTcpAddress, which limits the arbitrary-memory primitive by rejecting addresses outside registered buffers. However, the same report explicitly emphasizes that peer authentication is still missing on the TCP port, leaving the channel exposed to unauthenticated clients.

More importantly, separate disclosures demonstrate that 0.3.13 and even 0.3.13.post1 remain vulnerable to metadata server manipulation and memory exhaustion. The HTTP metadata server issue was still present in the main branch when the researcher reported it, and the handshake daemon flaw was confirmed against 0.3.13.post1. Security teams should therefore not treat 0.3.13 as a green light. Instead, they should track the maintainer's GitHub repository for subsequent releases that address the entire cluster, test patches in staging environments and verify that both the data plane and control plane are protected before declaring the environment safe (GitHub Issue #4441; GitHub Issue #4444; GitHub Issue #4445).

What should enterprise security teams do now?

Given the severity and breadth of the Mooncake disclosure cluster, CISOs and infrastructure owners should treat this as an active incident prevention exercise.

  • Inventory AI serving infrastructure. Mooncake may be deployed directly or embedded through platforms such as vLLM. Search software bills of materials, container images and running processes for mooncake-transfer-engine or related packages. An AI governance dashboard can help maintain visibility into which model serving stacks rely on Mooncake components.
  • Confirm network exposure. Verify that Mooncake TCP ports in the 15000-17000 range and any HTTP metadata interfaces are not reachable from untrusted networks, the public internet or general corporate VLANs. Where possible, bind these services to localhost or place them inside an authenticated service mesh with mutual TLS and strong identity.
  • Patch with verification. Upgrade to the latest available version beyond 0.3.13.post1 if the maintainers have released subsequent fixes, but validate that the update addresses the metadata server and handshake daemon issues as well as the TCP transport. Do not assume that 0.3.13 is a complete security resolution.
  • Segment and monitor. Apply zero-trust network policies so that only explicitly authorized inference nodes can reach Mooncake data and metadata ports. Monitor for anomalous inbound connections, unexpected memory growth in handshake daemons and unauthorized /metadata requests.
  • Rotate secrets and review logs. If Mooncake interfaces were exposed, treat the resident memory of affected processes as potentially compromised. Rotate any secrets that may have been present, and review access logs for the metadata server for unauthorized GET, PUT or DELETE activity.

The bigger picture: AI infrastructure is critical infrastructure

This disclosure is a reminder that the AI attack surface extends far beyond the model weights themselves. Data plane components that shuttle prompts, KV caches and secrets between GPUs and inference nodes are now as sensitive as the databases they resemble. A vulnerability in a transfer engine can be inherited silently when that engine is bundled into larger serving frameworks, creating supply chain risk that standard application scanning may miss.

Governance programs must therefore expand their scope from model cards and prompt policies to the network-facing infrastructure that serves those models. Organizations using Walled AI to govern employee AI use across browsers, desktop apps and APIs should ensure that the underlying serving infrastructure is included in asset inventories and vulnerability management scopes. The Mooncake cluster shows that an unauthenticated network interface deep in the AI stack can undo the compliance and data protection investments made at the application layer.

Frequently asked questions

What is Mooncake and why is it vulnerable?

Mooncake is an open-source data transfer engine that moves tensors, prompts and KV-cache data for large language model serving. It is vulnerable because its TCP transfer engine in versions before 0.3.13 accepted arbitrary memory addresses from the network without authentication or bounds checking, allowing remote attackers to read or write process memory directly.

Which Mooncake versions are affected by CVE-2026-103764?

CVE-2026-103764 affects Mooncake versions before 0.3.13. Version 0.3.13 added address validation that limits the arbitrary memory primitive, but it does not add peer authentication to the TCP port. Later reports also confirm that versions 0.3.13 and 0.3.13.post1 remain vulnerable to separate metadata server and denial-of-service flaws.

Can attackers exploit Mooncake without authentication?

Yes. The core TCP transfer flaw and the HTTP metadata server issue both allow unauthenticated remote access. An attacker on the network can send crafted requests to read or write memory, expose segment descriptors or exhaust daemon memory without providing credentials, making network segmentation and access control essential defenses.

What should I do if Mooncake is running in my environment?

Immediately inventory where Mooncake or mooncake-transfer-engine is deployed, including embedded use inside vLLM or other serving stacks. Ensure TCP ports 15000-17000 and any metadata interfaces are not exposed to untrusted networks. Upgrade to the latest patched release, verify that both data plane and control plane fixes are applied, and rotate any potentially exposed secrets.

Sources

Part of our daily Briefing Analysis, built on the AI Governance & Security Briefing.

MooncakeCVE-2026-103764AI infrastructure securityLLM servingunauthenticated vulnerabilityAI governance
Share LinkedIn X

Get audit-ready before August 2026.

See how WalledAI operationalises Articles 9–15 and Article 50 obligations as a runtime governance layer.