Skip to main content
Vulnerability Database/CVE-2026-96764

CVE-2026-96764: Mooncake kvcache-ai DOS Vulnerability

CVE-2026-96764 is a denial of service vulnerability in Mooncake kvcache-ai affecting versions up to 0.3.12 and 0.3.14-rc1. Attackers can exploit resource allocation flaws via regular expressions. This article covers technical details, affected versions, impact assessment, and mitigation strategies.

Updated:

CVE-2026-96764 Overview

CVE-2026-96764 is a resource allocation vulnerability in kvcache-ai mooncake versions up to 0.3.12 and 0.3.14-rc1. The flaw resides in the MasterService::GetReplicaListByRegex function within the Regular Expression Handler component. Attackers can send crafted requests remotely to trigger uncontrolled resource allocation, classified under [CWE-400]. A public exploit is available, and the vendor did not respond to coordinated disclosure attempts.

Critical Impact

Authenticated remote attackers can trigger resource exhaustion in the mooncake master service by submitting crafted regular expression inputs to GetReplicaListByRegex, degrading service availability.

Affected Products

  • kvcache-ai mooncake up to version 0.3.12
  • kvcache-ai mooncake 0.3.14-rc1
  • Deployments exposing the MasterService interface to untrusted callers

Discovery Timeline

  • 2026-09-24 - CVE CVE-2026-96764 published to NVD
  • 2026-09-24 - Last updated in NVD database

Technical Details for CVE-2026-96764

Vulnerability Analysis

The vulnerability affects MasterService::GetReplicaListByRegex, a handler that accepts a regular expression and matches it against replica identifiers. The function processes caller-supplied regex input without enforcing bounds on evaluation cost or returned result set size. Attackers manipulate this input to force the server into resource-intensive matching operations or large allocations.

Mooncake serves as a KVCache-centric disaggregated serving layer for large language model workloads, so disruption of its master service can stall inference pipelines. The attack requires network access and low privileges, with no user interaction. Impact is limited to availability; confidentiality and integrity are not affected.

Root Cause

The root cause is uncontrolled resource consumption [CWE-400] in the regex matching path. The GetReplicaListByRegex method does not constrain pattern complexity, matching time, or memory used to assemble replica lists. Pathological regular expressions and broad patterns can therefore drive CPU and memory usage disproportionately to request size.

Attack Vector

An authenticated remote attacker sends a request invoking GetReplicaListByRegex with a crafted pattern. The pattern either triggers catastrophic backtracking or matches a large portion of the replica namespace, forcing allocation of oversized data structures. Repeated invocations amplify the effect, degrading service for other tenants. A proof of concept is published publicly as a GitHub Gist code snippet.

No verified exploit code is reproduced here. Refer to the VulDB entry for CVE-2026-96764 for additional technical context.

Detection Methods for CVE-2026-96764

Indicators of Compromise

  • Sustained CPU or memory spikes in mooncake master service processes without a corresponding increase in legitimate client traffic
  • Repeated inbound RPC calls targeting MasterService::GetReplicaListByRegex from a narrow set of client identities
  • Request payloads containing nested quantifiers or wildcard-heavy regular expressions submitted to the master service

Detection Strategies

  • Instrument the GetReplicaListByRegex code path to log pattern length, execution time, and result set size for anomaly analysis
  • Baseline typical regex query shapes per tenant and alert on outliers in evaluation latency or returned replica counts
  • Correlate master service resource telemetry with RPC audit logs to attribute exhaustion events to specific callers

Monitoring Recommendations

  • Forward mooncake master service metrics and access logs to a centralized analytics pipeline for rate and pattern analysis
  • Track per-client request rates against the regex handler and alert when thresholds are exceeded
  • Monitor service restart counts and out-of-memory events on hosts running the mooncake master component

How to Mitigate CVE-2026-96764

Immediate Actions Required

  • Restrict network reachability of the mooncake master service to trusted inference workers and operators only
  • Revoke or rotate credentials for any accounts that do not require regex-based replica queries
  • Apply rate limiting to RPC endpoints that invoke GetReplicaListByRegex to cap per-caller request volume
  • Review published artifacts on the VulDB vulnerability page for additional guidance

Patch Information

The vendor was contacted but did not respond, and no official fix is referenced in the advisory data. Operators running kvcache-ai mooncake 0.3.12 or 0.3.14-rc1 should monitor the upstream project for a patched release and consider pinning deployments to validated builds. Until a fix is published, apply the workarounds below.

Workarounds

  • Place an authenticating proxy in front of the master service that validates regex inputs and rejects patterns exceeding length or complexity thresholds
  • Enforce timeouts and memory limits on regex evaluation at the process or container level using cgroups or runtime quotas
  • Disable or wrap the GetReplicaListByRegex endpoint if replica lookup by exact identifier is sufficient for the deployment
  • Isolate master service instances per tenant to contain the blast radius of availability degradation
bash
# Example container resource constraints to limit DoS impact
docker run --cpus="2.0" --memory="4g" --memory-swap="4g" \
  --name mooncake-master \
  --network mooncake-internal \
  kvcache-ai/mooncake:pinned

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.