I Hacked an MCP Server With No Login, Then Fixed It
PixelKart's support assistant gets its tools from an MCP (Model Context Protocol) server: get_order, refund_order and export_customers. The assistant only ever needs get_order. But the MCP server has no authentication.
PixelKart's support assistant gets its tools from an MCP (Model Context Protocol) server: get_order, refund_order and export_customers. The assistant only ever needs get_order.
But the MCP server has no authentication. So the attacker skips the chatbot entirely and talks to the server directly:
-
tools/listshows every tool -
export_customersdumps every customer's details -
refund_orderpays out βΉ99,999 on an order worth βΉ3,998
No prompt injection, no AI trickery. Just an open API with a new name.
Why it worked
- No authentication on the MCP endpoint. Anyone who could reach it could use it.
- One all-powerful identity. Every call ran with full access to orders and customers.
-
No input validation.
refund_orderaccepted any amount for any order.
MCP doesn't change the basics: a tool server is an API, and APIs need authentication, authorisation and validation.
The fix, in code
TOKENS = {ASSISTANT_TOKEN: {"client": "support-assistant", "scopes": {"orders:read"}}}
TOOL_SCOPE = {"get_order": "orders:read", "refund_order": "orders:refund",
"export_customers": "customers:export"}
def authorize(token, tool, args):
# 1. Every client needs a token.
if token not in TOKENS:
return False, "401 Unauthorized: missing or invalid token"
# 2. Each token only gets the scopes its job needs.
if tool and TOOL_SCOPE.get(tool) not in TOKENS[token]["scopes"]:
return False, f"403 Forbidden: this token can't call {tool}"
# 3. Validate inputs on the server, even for allowed tools.
if tool == "refund_order":
o = ORDERS.get(args.get("order_id"))
if not o or not 0 < float(args.get("amount", 0)) <= o["total"]:
return False, "403 Forbidden: refund amount not allowed"
return True, "ok"
tools/list also only shows what the token may call. Same attack, run again: 401 without a token, and even with the assistant's token (as if it leaked), export and refund are 403.
Run the lab yourself
git clone https://github.com/usinha79/hack-it-then-fix-it
cd hack-it-then-fix-it/ep05-mcp-server-auth
pip install -r requirements.txt
python vulnerable/app.py # then: python attack.py
export ASSISTANT_TOKEN=lab-assistant-7f3a91c2
python fixed/app.py # then: ./attack.sh
Full lab: github.com/usinha79/hack-it-then-fix-it
Take-home checklist
- Never expose an MCP server without authentication, even "only internally".
- Give each client a token scoped to the tools it needs.
- Validate every tool argument on the server.
- Log every MCP request with the client identity, and alert on unknown clients.
- Don't expose bulk-export tools to an assistant at all.
Hack It, Then Fix It is a series where I build a small AI app with a real flaw, break it, fix it in code, and run the same attack again to prove the fix holds. Everything shown uses fake data on my own machine. Only test systems you own or are authorised to test.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.