[web_server] Add CORS origin checking with allowed_origins (#17530)

THREAT_MODEL.md

@@ -92,18 +92,24 @@ is choosing an open control surface, in the same way that running native OTA
 without a password leaves OTA open. The API is documented and is meant to be
 called by other devices, scripts, and pages.
 
-The device performs no CSRF token, `Origin`, or `Referer` validation and returns
-a permissive CORS policy. Cross-origin requests are handled the same as any other
-network request, including requests a browser is induced to make by a page the
-operator visits (the "confused deputy", or CSRF, pattern). The following are
-therefore **not** vulnerabilities in this repository:
+As defense-in-depth, the web server checks the `Origin` header on browser requests
+to its entity control and state endpoints: a request whose `Origin` does not match
+the address the device is served on is rejected, and the `allowed_origins` option
+widens that list. This blocks the common "confused deputy" (CSRF) case where a page
+the operator visits drives the device through their browser. It is **not** an
+authentication boundary: it only constrains browsers. Any client that omits the
+`Origin` header — `curl`, scripts, or other non-browser callers on the same
+network — reaches every endpoint exactly as before. The check also does not cover
+the web OTA `/update` endpoint. The device performs no CSRF-token or `Referer`
+validation. The following are therefore **not** vulnerabilities in this repository:
 
-- Cross-origin or CSRF requests to the control endpoints (for example, a page the
-  operator opens toggling a switch), whether or not `web_server` `auth:` is set.
-- Cross-origin reads of device state permitted by the CORS policy.
-- Cross-origin firmware upload through the web OTA endpoint (`/update`) when web
-  OTA is enabled without `web_server` `auth:`. This is the same exposure as
-  running OTA without a password.
+- Requests without an `Origin` header (for example `curl`) reaching the control
+  endpoints, whether or not `web_server` `auth:` is set.
+- Requests from an origin the operator added to `allowed_origins`.
+- Cross-origin or CSRF firmware upload through the web OTA endpoint (`/update`) when
+  web OTA is enabled without `web_server` `auth:`. The `/update` endpoint is not
+  covered by the `Origin` check; this is the same exposure as running OTA without a
+  password.
 
 The supported defenses are `web_server` `auth:`, protecting OTA (a web password or
 a native OTA password), and keeping devices on a trusted, segmented network. See
@@ -113,9 +119,7 @@ What remains in scope is bypassing `web_server` `auth:` when it *is* configured,
 and any memory-safety or protocol bug in the server reachable without credentials.
 
 This section documents the current design and scope; it is not a judgment that the
-design is optimal or that it will not change. Optional hardening (for example an
-origin allowlist or opt-in CSRF checks) is welcome as a normal enhancement PR,
-framed as defense-in-depth rather than a security fix.
+design is optimal or that it will not change.
 
 ## Explicitly out of scope
 
@@ -124,9 +128,10 @@ framed as defense-in-depth rather than a security fix.
 - Operator-supplied hostile YAML (covered above — config authoring is trusted).
 - Attacks that require an already-authenticated device peer (someone who already
   holds the API key / OTA / web credentials).
-- Cross-site (CSRF), cross-origin, or CORS behavior of the device web server and
-  its web OTA endpoint. The web server is an open HTTP API by design (see above);
-  gate it with `web_server` `auth:` and network isolation.
+- Access to the device web server or its web OTA endpoint by non-browser clients
+  (those that send no `Origin` header). The web server is an open HTTP API by
+  design (see above); browser cross-origin requests are blocked by default, but the
+  real controls are `web_server` `auth:` and network isolation.
 - Anything in the dashboard / device-builder — report that in its own repository
   (linked at the top).
 - Deployments where the operator removed protections or exposed credentials. See

esphome/components/web_server/__init__.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import gzip
 import logging
+import re
 
 import esphome.codegen as cg
 from esphome.components import web_server_base
@@ -46,6 +47,7 @@ AUTO_LOAD = ["json", "web_server_base"]
 CONF_SORTING_GROUP_ID = "sorting_group_id"
 CONF_SORTING_GROUPS = "sorting_groups"
 CONF_SORTING_WEIGHT = "sorting_weight"
+CONF_ALLOWED_ORIGINS = "allowed_origins"
 
 
 web_server_ns = cg.esphome_ns.namespace("web_server")
@@ -104,6 +106,41 @@ def validate_ota(config: ConfigType) -> ConfigType:
     return config
 
 
+# An Origin header is always "scheme://host[:port]" with no path or trailing slash.
+_ORIGIN_RE = re.compile(r"^[a-zA-Z][a-zA-Z0-9+.-]*://[^/\s]+$")
+
+
+def validate_origin(value: str) -> str:
+    # "*" is the wildcard that allows any origin.
+    if value == "*":
+        return value
+    value = cv.string_strict(value)
+    if not _ORIGIN_RE.match(value):
+        raise cv.Invalid(
+            f"'{value}' is not a valid origin. An origin must be 'scheme://host[:port]' with no "
+            f"path or trailing slash (e.g. 'https://example.com'), or '*' to allow any origin."
+        )
+    # Browsers send the scheme and host lowercased in the Origin header, so normalize to match.
+    return value.lower()
+
+
+def validate_private_network_access(config: ConfigType) -> ConfigType:
+    # PNA preflights are always cross-origin, so they can only be authorized against the
+    # allowed_origins list. Enabling PNA without any origins would deny every PNA request.
+    if (
+        config[CONF_ENABLE_PRIVATE_NETWORK_ACCESS]
+        and config.get(CONF_ALLOWED_ORIGINS) is None
+    ):
+        raise cv.Invalid(
+            f"'{CONF_ALLOWED_ORIGINS}' must be set when "
+            f"'{CONF_ENABLE_PRIVATE_NETWORK_ACCESS}' is enabled. List each origin that is "
+            f"allowed to reach the device (e.g. 'https://example.com'). '*' allows any origin "
+            f"but is not recommended.",
+            path=[CONF_ENABLE_PRIVATE_NETWORK_ACCESS],
+        )
+    return config
+
+
 def validate_sorting_groups(config: ConfigType) -> ConfigType:
     if CONF_SORTING_GROUPS in config and config[CONF_VERSION] != 3:
         raise cv.Invalid(
@@ -201,7 +238,10 @@ CONFIG_SCHEMA = cv.All(
             cv.Optional(CONF_CSS_INCLUDE): cv.file_,
             cv.Optional(CONF_JS_URL): cv.string,
             cv.Optional(CONF_JS_INCLUDE): cv.file_,
-            cv.Optional(CONF_ENABLE_PRIVATE_NETWORK_ACCESS, default=True): cv.boolean,
+            cv.Optional(CONF_ENABLE_PRIVATE_NETWORK_ACCESS, default=False): cv.boolean,
+            cv.Optional(CONF_ALLOWED_ORIGINS): cv.All(
+                cv.ensure_list(validate_origin), cv.Length(min=1)
+            ),
             cv.Optional(CONF_AUTH): cv.Schema(
                 {
                     cv.Required(CONF_USERNAME): cv.All(
@@ -238,6 +278,7 @@ CONFIG_SCHEMA = cv.All(
     validate_local,
     validate_sorting_groups,
     validate_ota,
+    validate_private_network_access,
     _consume_web_server_sockets,
 )
 
@@ -334,6 +375,9 @@ async def to_code(config):
         request_log_listener()  # Request a log listener slot for web server log streaming
     if config[CONF_ENABLE_PRIVATE_NETWORK_ACCESS]:
         cg.add_define("USE_WEBSERVER_PRIVATE_NETWORK_ACCESS")
+    if (allowed_origins := config.get(CONF_ALLOWED_ORIGINS)) is not None:
+        cg.add_define("USE_WEBSERVER_ALLOWED_ORIGINS")
+        cg.add(var.set_allowed_origins(allowed_origins))
     if CONF_AUTH in config:
         cg.add_define("USE_WEBSERVER_AUTH")
         cg.add(paren.set_auth_username(config[CONF_AUTH][CONF_USERNAME]))

esphome/components/web_server/web_server.cpp

@@ -456,9 +456,58 @@ void WebServer::handle_index_request(AsyncWebServerRequest *request) {
 }
 #endif
 
+// Read a request header value portably across the Arduino and ESP-IDF web servers.
+// Returns an empty string when the header is absent (only allocates when a value is present).
+static std::string get_request_header(AsyncWebServerRequest *request, const char *name) {
+#ifdef USE_ESP32
+  // ESP32 (Arduino and ESP-IDF) uses the web_server_idf backend.
+  optional<std::string> value = request->get_header(name);
+  return value.has_value() ? std::move(*value) : std::string();
+#else
+  // ESP8266, RP2040 and LibreTiny use the Arduino ESPAsyncWebServer backend.
+  const AsyncWebHeader *header = request->getHeader(name);
+  return header != nullptr ? std::string(header->value().c_str()) : std::string();
+#endif
+}
+
+bool WebServer::is_request_origin_allowed_(AsyncWebServerRequest *request, const std::string &origin) {
+  // No Origin header: not a browser cross-origin request (e.g. curl, native API client). Allow.
+  if (origin.empty())
+    return true;
+
+  // Same-origin: the Origin authority (scheme stripped) matches the Host the request was sent to.
+  // This covers the device's own IP, mDNS name, or DNS name without knowing any at compile time.
+  const size_t scheme_sep = origin.find("://");
+  if (scheme_sep != std::string::npos) {
+    const std::string host = get_request_header(request, "Host");
+    if (!host.empty() && origin.compare(scheme_sep + 3, std::string::npos, host) == 0)
+      return true;
+  }
+
+#ifdef USE_WEBSERVER_ALLOWED_ORIGINS
+  // Otherwise the origin must be explicitly allowed via configuration.
+  for (const char *allowed_origin : this->allowed_origins_) {
+    // A single "*" entry allows any origin.
+    if (allowed_origin[0] == '*' && allowed_origin[1] == '\0')
+      return true;
+    if (origin == allowed_origin)
+      return true;
+  }
+#endif
+  return false;
+}
+
 #ifdef USE_WEBSERVER_PRIVATE_NETWORK_ACCESS
 void WebServer::handle_pna_cors_request(AsyncWebServerRequest *request) {
+  const std::string origin = get_request_header(request, "Origin");
+  if (!this->is_request_origin_allowed_(request, origin)) {
+    request->send(403);
+    return;
+  }
+
   AsyncWebServerResponse *response = request->beginResponse(200, ESPHOME_F(""));
+  // Echo the specific origin back so the response is valid even when auth (credentials) is enabled.
+  response->addHeader(ESPHOME_F("Access-Control-Allow-Origin"), origin.empty() ? "*" : origin.c_str());
   response->addHeader(ESPHOME_F("Access-Control-Allow-Private-Network"), ESPHOME_F("true"));
   response->addHeader(ESPHOME_F("Private-Network-Access-Name"), App.get_name().c_str());
   char mac_s[18];
@@ -2448,6 +2497,21 @@ void WebServer::handleRequest(AsyncWebServerRequest *request) {
     return;
   }
 
+#ifdef USE_WEBSERVER_PRIVATE_NETWORK_ACCESS
+  // Private Network Access preflight carries a cross-origin Origin by design; its handler does the
+  // origin check itself, so let it run before the general enforcement below.
+  if (request->method() == HTTP_OPTIONS && request->hasHeader(ESPHOME_F("Access-Control-Request-Private-Network"))) {
+    this->handle_pna_cors_request(request);
+    return;
+  }
+#endif
+
+  // Reject cross-origin browser requests unless the origin is explicitly allowed.
+  if (!this->is_request_origin_allowed_(request, get_request_header(request, "Origin"))) {
+    request->send(403);
+    return;
+  }
+
 #if !defined(USE_ESP32) && defined(USE_ARDUINO)
   if (url == ESPHOME_F("/events")) {
     this->events_.add_new_client(this, request);
@@ -2469,13 +2533,6 @@ void WebServer::handleRequest(AsyncWebServerRequest *request) {
   }
 #endif
 
-#ifdef USE_WEBSERVER_PRIVATE_NETWORK_ACCESS
-  if (request->method() == HTTP_OPTIONS && request->hasHeader(ESPHOME_F("Access-Control-Request-Private-Network"))) {
-    this->handle_pna_cors_request(request);
-    return;
-  }
-#endif
-
   // Parse URL for component routing
   // Pass HTTP method to disambiguate 3-segment URLs (GET=sub-device state, POST=main device action)
   UrlMatch match = match_url(url.c_str(), url.length(), false, request->method() == HTTP_POST);

esphome/components/web_server/web_server.h

@@ -242,6 +242,22 @@ class WebServer final : public Controller, public Component, public AsyncWebHand
    */
   void set_expose_log(bool expose_log) { this->expose_log_ = expose_log; }
 
+#ifdef USE_WEBSERVER_ALLOWED_ORIGINS
+  /** Set the origins that browsers are allowed to make cross-origin requests from.
+   *
+   * Requests without an `Origin` header (e.g. non-browser clients like curl or the native API)
+   * are always allowed. Requests whose `Origin` matches the address the device is served on
+   * (same-origin) are always allowed. Any other browser origin must appear in this list, or the
+   * request is rejected. A single "*" entry allows any origin. Each other entry must exactly match
+   * the requesting page's `Origin` header (e.g. "https://example.com").
+   *
+   * This list is also used to authorize Private Network Access requests when that feature is enabled.
+   *
+   * @param origins The list of allowed origins.
+   */
+  void set_allowed_origins(std::initializer_list<const char *> origins) { this->allowed_origins_ = origins; }
+#endif
+
   // ========== INTERNAL METHODS ==========
   // (In most use cases you won't need these)
   /// Setup the internal web server and register handlers.
@@ -593,6 +609,16 @@ class WebServer final : public Controller, public Component, public AsyncWebHand
   const char *js_include_{nullptr};
 #endif
   bool expose_log_{true};
+#ifdef USE_WEBSERVER_ALLOWED_ORIGINS
+  // Extra origins allowed to make cross-origin browser requests ("*" means any origin).
+  // Only compiled when allowed_origins is configured; same-origin is always allowed regardless.
+  FixedVector<const char *> allowed_origins_;
+#endif
+
+  /// Check whether the given request Origin is permitted. Same-origin (matching the Host the
+  /// request was sent to) and requests without an Origin header are always allowed; any other
+  /// origin must be listed in allowed_origins. The caller passes the already-read Origin header.
+  bool is_request_origin_allowed_(AsyncWebServerRequest *request, const std::string &origin);
 
  private:
 #ifdef USE_SENSOR

esphome/core/defines.h

@@ -302,6 +302,7 @@
 #define USE_WEBSERVER_PORT 80  // NOLINT
 #define USE_WEBSERVER_GZIP
 #define USE_WEBSERVER_SORTING
+#define USE_WEBSERVER_ALLOWED_ORIGINS
 #define WEB_SERVER_DEFAULT_HEADERS_COUNT 1
 #define USE_CAPTIVE_PORTAL_GZIP
 #define USE_WIFI_11KV_SUPPORT

tests/component_tests/web_server/test_private_network_access.py

@@ -0,0 +1,86 @@
+"""Tests for web_server Private Network Access / allowed_origins validation."""
+
+import pytest
+
+from esphome import config_validation as cv
+from esphome.components.web_server import (
+    CONF_ALLOWED_ORIGINS,
+    validate_origin,
+    validate_private_network_access,
+)
+from esphome.const import CONF_ENABLE_PRIVATE_NETWORK_ACCESS
+from esphome.types import ConfigType
+
+
+def test_pna_enabled_without_origins_fails() -> None:
+    """Enabling PNA without allowed_origins must fail validation."""
+    config: ConfigType = {CONF_ENABLE_PRIVATE_NETWORK_ACCESS: True}
+
+    with pytest.raises(cv.Invalid) as exc_info:
+        validate_private_network_access(config)
+
+    error_msg = str(exc_info.value)
+    assert CONF_ALLOWED_ORIGINS in error_msg
+    assert "must be set" in error_msg
+
+
+def test_pna_enabled_with_origins_passes() -> None:
+    """Enabling PNA with at least one allowed origin passes validation."""
+    config: ConfigType = {
+        CONF_ENABLE_PRIVATE_NETWORK_ACCESS: True,
+        CONF_ALLOWED_ORIGINS: ["https://app.esphome.io"],
+    }
+    assert validate_private_network_access(config) == config
+
+
+def test_origins_without_pna_passes() -> None:
+    """allowed_origins can be set without enabling PNA (they are independent)."""
+    config: ConfigType = {
+        CONF_ENABLE_PRIVATE_NETWORK_ACCESS: False,
+        CONF_ALLOWED_ORIGINS: ["https://app.esphome.io"],
+    }
+    assert validate_private_network_access(config) == config
+
+
+def test_pna_disabled_without_origins_passes() -> None:
+    """PNA disabled and no origins specified passes validation."""
+    config: ConfigType = {CONF_ENABLE_PRIVATE_NETWORK_ACCESS: False}
+    assert validate_private_network_access(config) == config
+
+
+def test_validate_origin_wildcard() -> None:
+    """The '*' wildcard is accepted as-is."""
+    assert validate_origin("*") == "*"
+
+
+@pytest.mark.parametrize(
+    "value",
+    [
+        "https://example.com",
+        "http://example.com:8080",
+        "https://192.168.1.5",
+    ],
+)
+def test_validate_origin_valid(value: str) -> None:
+    """Well-formed origins pass through unchanged."""
+    assert validate_origin(value) == value
+
+
+def test_validate_origin_lowercased() -> None:
+    """Scheme and host are normalized to lowercase to match the browser Origin header."""
+    assert validate_origin("HTTPS://App.Example.com") == "https://app.example.com"
+
+
+@pytest.mark.parametrize(
+    "value",
+    [
+        "https://example.com/",  # trailing slash
+        "https://example.com/path",  # path segment
+        "example.com",  # missing scheme
+        "",  # empty
+    ],
+)
+def test_validate_origin_invalid(value: str) -> None:
+    """Malformed origins are rejected at config time instead of silently 403ing."""
+    with pytest.raises(cv.Invalid, match="not a valid origin"):
+        validate_origin(value)

tests/components/web_server/common_v2.yaml

@@ -5,3 +5,6 @@ web_server:
   port: 8080
   version: 2
   compression: br
+  enable_private_network_access: true
+  allowed_origins:
+    - https://app.esphome.io

tests/components/web_server/common_v3.yaml

@@ -4,6 +4,9 @@ packages:
 web_server:
   port: 8080
   version: 3
+  # allowed_origins can be set independently of Private Network Access
+  allowed_origins:
+    - https://app.esphome.io
   sorting_groups:
     - id: sorting_group_1
       name: "Group 1 Diplayed Last"