PraisonAI ToolsMCPServer legacy SSE transport accepts attacker Host/Origin and exposes registered tools

{
  "affected": [
    {
      "database_specific": {
        "last_known_affected_version_range": "\u003c= 1.6.58"
      },
      "package": {
        "ecosystem": "PyPI",
        "name": "praisonaiagents"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0.6.0"
            },
            {
              "fixed": "1.6.59"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    },
    {
      "database_specific": {
        "last_known_affected_version_range": "\u003c= 4.6.58"
      },
      "package": {
        "ecosystem": "PyPI",
        "name": "praisonai"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "3.10.0"
            },
            {
              "fixed": "4.6.59"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "aliases": [],
  "database_specific": {
    "cwe_ids": [
      "CWE-306",
      "CWE-346",
      "CWE-862"
    ],
    "github_reviewed": true,
    "github_reviewed_at": "2026-06-18T13:52:40Z",
    "nvd_published_at": null,
    "severity": "HIGH"
  },
  "details": "# PraisonAI ToolsMCPServer legacy SSE transport accepts attacker Host/Origin and exposes registered tools\n\n## Summary\n\n`praisonaiagents.mcp.ToolsMCPServer.run_sse()` builds a Starlette MCP\nHTTP+SSE server around `mcp.server.sse.SseServerTransport`. The server exposes\n`/sse` and `/messages/`, but it does not validate `Origin`, does not validate\n`Host`, and does not require any authentication.\n\nThis is reachable through supported PraisonAI code paths that wrap configured\nMCP server tools and re-expose them over legacy SSE:\n\n- `praisonai mcp run \u003cname\u003e --transport sse`\n- `praisonai serve mcp --name \u003cname\u003e --transport sse`\n- direct use of `ToolsMCPServer(...).run_sse(...)` or\n  `launch_tools_mcp_server(..., transport=\"sse\")`\n\nA malicious website can use DNS rebinding against a local or internal\nPraisonAI SSE MCP server and send requests with attacker-controlled `Host` and\n`Origin` headers. The local PoV binds only to `127.0.0.1`, sends an attacker\n`Host` and `Origin`, lists the registered tool, and invokes it successfully.\n\nThe same attacker `Origin` is rejected by PraisonAI\u0027s current Streamable HTTP\ntransport with HTTP 403. The vulnerability is therefore a sibling transport\nguard gap in the legacy SSE wrapper, not intended behavior.\n\n## Affected product\n\n- Repository: `MervinPraison/PraisonAI`\n- Packages:\n  - `praisonaiagents`\n  - `praisonai`\n- Primary component:\n  `src/praisonai-agents/praisonaiagents/mcp/mcp_server.py`\n- CLI wrappers:\n  - `src/praisonai/praisonai/cli/commands/mcp.py`\n  - `src/praisonai/praisonai/cli/commands/serve.py`\n- Latest verified release/current head:\n  - `praisonaiagents 1.6.58`\n  - `PraisonAI 4.6.58`\n  - repo head `1ad58ca02975ff1398efeda694ea2ab78f20cf3e`\n\nSuggested affected ranges:\n\n- `praisonaiagents \u003e= 0.6.0, \u003c= 1.6.58`\n- `praisonai \u003e= 3.10.0, \u003c= 4.6.58`\n\nNo fixed version is known at submission time.\n\nConfirmed source sweep:\n\n```text\nv3.0.0   ToolsMCPServer.run_sse helper present, no Origin/Host/auth checks\nv3.10.0  praisonai mcp run --transport sse wraps configured tools into helper\nv3.12.3  praisonai serve mcp --name --transport sse wraps configured tools\nv4.0.0   same vulnerable helper and CLI wrapping paths\nv4.4.12  same vulnerable helper and CLI wrapping paths\nv4.5.0   same vulnerable helper and CLI wrapping paths\nv4.5.56  same vulnerable helper and CLI wrapping paths\nv4.5.139 same vulnerable helper and CLI wrapping paths\nv4.6.57  same vulnerable helper and CLI wrapping paths\nv4.6.58  same vulnerable helper and dynamic PoV succeeds\n```\n\n## Impact\n\nIf a PraisonAI user starts a local or internal legacy SSE MCP server with\nregistered tools, an attacker who gets that user to visit a malicious website\ncan use DNS rebinding to interact with the SSE server through the browser. The\nattacker can discover exposed tools and invoke them as the local user.\n\nImpact depends on the configured tools. In realistic PraisonAI MCP deployments,\nregistered tools may access local files, repositories, issue trackers, cloud\nAPIs, internal services, or other automation targets. This can lead to\nconfidentiality, integrity, and availability impact for the resources reachable\nby the exposed tools.\n\nThe PoV is local-only and harmless. It exposes one marker tool that writes a\ncanary string to a temporary directory.\n\n## Root cause\n\nCurrent `ToolsMCPServer.run_sse()` constructs a Starlette app directly:\n\n```python\nsse_path = \"/sse\"\nmessages_path = \"/messages/\"\nsse_transport = SseServerTransport(messages_path)\n\nasync def handle_sse(request: Request):\n    async with sse_transport.connect_sse(\n        request.scope, request.receive, request._send\n    ) as (read_stream, write_stream):\n        await mcp._mcp_server.run(\n            read_stream,\n            write_stream,\n            mcp._mcp_server.create_initialization_options()\n        )\n\napp = Starlette(\n    debug=self._debug,\n    routes=[\n        Route(sse_path, endpoint=handle_sse),\n        Mount(messages_path, app=sse_transport.handle_post_message),\n    ]\n)\n\nuvicorn.run(app, host=host, port=port)\n```\n\nThere is no middleware or route-level check for:\n\n- `Origin`\n- `Host`\n- `Authorization`\n- API key\n- allowed origins / allowed hosts\n\nThe configured CLI wrapper exposes this path:\n\n```python\nfrom praisonaiagents.mcp import MCP, ToolsMCPServer\ncmd_string = \" \".join(cmd)\nmcp = MCP(cmd_string, timeout=60, env=server.env or {})\ntools = mcp.get_tools()\nmcp_server = ToolsMCPServer(name=name, tools=tools)\nmcp_server.run_sse(host=host, port=port)\n```\n\nBy contrast, the current Streamable HTTP transport validates `Origin` and\nreturns HTTP 403 for an invalid origin:\n\n```python\norigin = request.headers.get(\"Origin\")\nif not self._validate_origin(origin):\n    return JSONResponse(..., status_code=403)\n```\n\n## Local-only PoV\n\nRun from the harness checkout:\n\n```bash\nuv run --with mcp --with starlette --with uvicorn --with httpx --with anyio \\\n  python submission-bundle/praisonai-prai-cand-015-mcp-sse-host-origin-bypass/poc/pov_prai_cand_015_sse_mcp_host_origin_bypass.py \\\n  --repo-src artifacts/repos/praisonai-v4.6.58/src\n```\n\nObserved current-head result:\n\n```json\n{\n  \"candidate\": \"PRAI-CAND-015\",\n  \"http_stream_control\": {\n    \"attacker_origin\": \"http://attacker.example.test\",\n    \"rejects_attacker_origin\": true,\n    \"status_code\": 403,\n    \"transport\": \"current_http_stream\"\n  },\n  \"source_checks\": {\n    \"has_auth_check\": false,\n    \"has_host_check\": false,\n    \"has_origin_check\": false,\n    \"has_sse_transport\": true,\n    \"route_count\": 2\n  },\n  \"sse_probe\": {\n    \"attacker_headers\": {\n      \"Host\": \"attacker.example.test:62380\",\n      \"Origin\": \"http://attacker.example.test:62380\"\n    },\n    \"bind_host\": \"127.0.0.1\",\n    \"marker_value\": \"executed-from-attacker-origin\",\n    \"marker_written\": true,\n    \"server_started\": true,\n    \"tool_call_content\": [\n      \"recorded:executed-from-attacker-origin\"\n    ],\n    \"tool_call_error\": false,\n    \"tool_names\": [\n      \"record_marker\"\n    ],\n    \"vulnerable\": true\n  },\n  \"vulnerable\": true\n}\n```\n\nThe PoV:\n\n1. imports the current `ToolsMCPServer`;\n2. registers one marker tool;\n3. monkey-patches `uvicorn.run` only to capture the exact Starlette app created\n   by `run_sse()`;\n4. starts that app on `127.0.0.1`;\n5. connects to `/sse` with attacker-controlled `Host` and `Origin`;\n6. lists tools and calls the marker tool;\n7. runs a control against PraisonAI\u0027s current Streamable HTTP transport and\n   confirms the same attacker `Origin` is rejected with HTTP 403.\n\n## Why this is not intended behavior\n\nThis is not only a trust-model disagreement.\n\nPraisonAI\u0027s MCP documentation describes Streamable HTTP, WebSocket, and legacy\nSSE as supported MCP transport mechanisms. The same documentation says the MCP\nmodule\u0027s security properties include origin validation, authentication headers,\nand secure session IDs. The transport guide also has a dedicated security\nsection for origin validation as DNS rebinding prevention and authentication.\n\nThe official MCP specification warns that HTTP transports need origin\nvalidation to prevent DNS rebinding, should bind locally for local servers, and\nshould implement authentication. It also says that without those protections,\nremote websites can interact with local MCP servers.\n\nThe upstream MCP Python SDK advisory `GHSA-9h52-p55h-vw2f` / `CVE-2025-66416`\nclassifies unauthenticated localhost HTTP/SSE MCP servers without DNS rebinding\nprotection as a High severity issue because malicious websites can invoke tools\nor access resources exposed by the local MCP server. That advisory also says\ncustom low-level `SseServerTransport` configurations should explicitly configure\ntransport security settings when running unauthenticated localhost servers.\n\nPraisonAI\u0027s current Streamable HTTP implementation already enforces an Origin\nguard and rejects the exact attacker Origin used in the PoV. The issue is that\nthe legacy SSE sibling path lacks the same boundary.\n\n\n## Suggested severity\n\nSuggested severity: High.\n\nRationale:\n\n- `AV`: the attack uses browser-origin HTTP requests to a local/internal\n  service.\n- `AC`: practical exploitation requires DNS rebinding or equivalent browser\n  origin setup.\n- `PR`: no PraisonAI credentials are required by the SSE server.\n- `UR`: the user must visit an attacker-controlled page.\n- `S`: the vulnerable transport exposes tools that operate on resources\n  outside the HTTP transport itself.\n- `C/I/A`: exposed tools may read, mutate, or disrupt local/internal\n  resources depending on the configured MCP server.\n\n## Suggested fix\n\nBring legacy SSE server security in line with the current Streamable HTTP\ntransport, or disable the legacy SSE server path.\n\nRecommended changes:\n\n1. Add explicit allowed-origin and allowed-host validation to both `/sse` and\n   `/messages/`.\n2. Reject invalid `Origin` with HTTP 403 before opening the SSE stream or\n   accepting POST messages.\n3. Validate `Host` for local and internal deployments to mitigate DNS rebinding\n   even when browsers omit or vary `Origin`.\n4. Require authentication for all non-stdio MCP HTTP transports, including SSE.\n5. Add `--api-key`, `--allowed-origins`, and `--allowed-hosts` options to\n   `praisonai mcp run` and `praisonai serve mcp` when `--transport sse` is used.\n6. Where the installed MCP SDK supports it, configure the SDK transport-security\n   settings for low-level `SseServerTransport` usage instead of mounting it\n   without Host/Origin protection.\n7. Consider deprecating or disabling `--transport sse` server mode in favor of\n   the current Streamable HTTP implementation.\n8. Add regression tests proving that attacker `Host` and `Origin` values are\n   rejected on both `/sse` and `/messages/`, and that current Streamable HTTP and\n   legacy SSE enforce the same boundary.",
  "id": "GHSA-vmf9-xx9w-86wx",
  "modified": "2026-06-18T13:52:40Z",
  "published": "2026-06-18T13:52:40Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/MervinPraison/PraisonAI/security/advisories/GHSA-vmf9-xx9w-86wx"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/MervinPraison/PraisonAI"
    }
  ],
  "schema_version": "1.4.0",
  "severity": [
    {
      "score": "CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:C/C:H/I:H/A:H",
      "type": "CVSS_V3"
    }
  ],
  "summary": "PraisonAI ToolsMCPServer legacy SSE transport accepts attacker Host/Origin and exposes registered tools"
}