diff --git a/agent/anthropic_adapter.py b/agent/anthropic_adapter.py
index e93d068129..1782c72673 100644
--- a/agent/anthropic_adapter.py
+++ b/agent/anthropic_adapter.py
@@ -340,11 +340,13 @@ def _build_anthropic_client_with_bearer_hook(
     kwargs["http_client"] = build_bearer_http_client(token_provider, timeout=kwargs["timeout"])
     kwargs["auth_token"] = "entra-id-bearer-via-http-hook"
     headers = _beta_header(_common_betas_for_base_url(normalized_base_url, drop_context_1m_beta=drop_context_1m_beta))
-    return _new_sdk_client(sdk, kwargs, headers)
+    return _new_sdk_client(sdk, kwargs, headers, route=base_url)
 
 
-def _new_sdk_client(sdk, kwargs: Dict[str, Any], headers: Dict[str, str]):
+def _new_sdk_client(sdk, kwargs: Dict[str, Any], headers: Dict[str, str], route: str = None):
     """``sdk.Anthropic(**kwargs)`` with ``headers`` attached, sending exactly ONE credential.
+    ``route`` is the caller's un-normalized base_url (the ``/v1`` form ``custom_providers`` entries are
+    keyed by; ``kwargs["base_url"]`` has it stripped) for the per-provider ``extra_headers`` lookup.
 
     The SDK fills whichever of ``api_key`` / ``auth_token`` we left unset from ANTHROPIC_API_KEY /
     ANTHROPIC_AUTH_TOKEN in the environment (both loaded from ~/.hermes/.env) and then sends dual
@@ -357,11 +359,28 @@ def _new_sdk_client(sdk, kwargs: Dict[str, Any], headers: Dict[str, str]):
         merged["Authorization"] = sdk.Omit()
     elif "auth_token" in kwargs and "api_key" not in kwargs:
         merged["X-Api-Key"] = sdk.Omit()
+    # Per-provider ``custom_providers[].extra_headers`` last: the most specific config level wins
+    # over the SDK User-Agent and the attribution/beta sets above, on every builder path (init,
+    # /model switch, rebuild, auxiliary) — the OpenAI-wire clients already do this (#24293, #9721).
+    merged.update(_custom_provider_extra_headers(route or kwargs.get("base_url")))
     if merged:
         kwargs["default_headers"] = merged
     return sdk.Anthropic(**kwargs)
 
 
+def _custom_provider_extra_headers(base_url) -> Dict[str, str]:
+    """``extra_headers`` of the ``custom_providers`` entry routed at *base_url*, else ``{}``.
+    SECURITY: values routinely carry credentials (Cloudflare Access tokens) — never log them."""
+    if not base_url:
+        return {}
+    try:
+        from hermes_cli.config import get_custom_provider_extra_headers
+        return get_custom_provider_extra_headers(str(base_url))
+    except Exception:
+        logger.debug("custom-provider extra_headers skipped for Anthropic client", exc_info=True)
+        return {}
+
+
 def _auth_style(api_key, base_url, normalized_base_url) -> str:
     """Order-sensitive endpoint/key classification for :func:`build_anthropic_client`. ``kimi``:
     Kimi's /coding endpoint 403s without a User-Agent (the Kimi team asked for proper attribution).
@@ -410,7 +429,7 @@ def build_anthropic_client(api_key, base_url: str = None, timeout: float = None,
         # get these from profile.default_headers, but this route never sees the profile.
         for k, v in _attribution_headers().items():
             headers.setdefault(k, v)
-    return _new_sdk_client(sdk, kwargs, headers)
+    return _new_sdk_client(sdk, kwargs, headers, route=base_url)
 
 
 def build_anthropic_bedrock_client(region: str):
diff --git a/agent/error_classifier.py b/agent/error_classifier.py
index f951a7458e..7e190cf900 100644
--- a/agent/error_classifier.py
+++ b/agent/error_classifier.py
@@ -38,6 +38,7 @@ class FailoverReason(enum.Enum):
     billing = "billing"                  # 402 or confirmed credit exhaustion — rotate immediately
     rate_limit = "rate_limit"            # 429 or quota-based throttling — backoff then rotate
     upstream_rate_limit = "upstream_rate_limit"  # Aggregator's upstream model 429 — fallback model, key is healthy
+    upstream_blocked = "upstream_blocked"  # 403 from a WAF/CDN/proxy in front of the provider — key is healthy, fallback
     overloaded = "overloaded"            # 503/529 — provider overloaded, backoff
     server_error = "server_error"        # 500/502 — internal server error, retry
     timeout = "timeout"                  # Connection/read timeout — rebuild client + retry
@@ -416,6 +417,17 @@ _SSL_TRANSIENT_PATTERNS = (
 )
 
 
+# A 403 body written by a WAF/CDN/proxy rather than the provider's API: Cloudflare's browser
+# challenge and block pages, plus the plain-text block relays return when they reject the SDK
+# User-Agent (#53099). Matched only on 403 (see ``_status_403``); a bare "access denied" or
+# "forbidden" stays auth because providers word real permission errors that way too.
+_UPSTREAM_BLOCKED_PATTERNS = (
+    "your request was blocked", "request blocked", "sorry, you have been blocked",
+    "enable javascript and cookies to continue", "cdn-cgi/challenge-platform", "cf-browser-verification",
+    "challenge-error-text", "__cf_chl", "cf-error-details", "attention required! | cloudflare",
+)
+
+
 # ── Verdicts and rule tables ────────────────────────────────────────────
 # A verdict is the ClassifiedError kwargs a stage decided on: ``reason`` plus
 # hint overrides (unlisted hints keep dataclass defaults). Rule tables are
@@ -438,6 +450,7 @@ _V_RATE_LIMIT = _v(_R.rate_limit, **_ROTATE_FALLBACK)
 _V_AUTH_ROTATE = _v(_R.auth, retryable=False, **_ROTATE_FALLBACK)
 _V_AUTH_FALLBACK = _v(_R.auth, **_ABORT_FALLBACK)
 _V_MODEL_NOT_FOUND = _v(_R.model_not_found, **_ABORT_FALLBACK)
+_V_UPSTREAM_BLOCKED = _v(_R.upstream_blocked, **_ABORT_FALLBACK)
 _V_CONTENT_BLOCKED = _v(_R.content_policy_blocked, **_ABORT_FALLBACK)
 # Another account in the same pool may hold the entitlement; the credential itself is healthy.
 _V_MODEL_ENTITLEMENT = _v(_R.model_entitlement, retryable=False, **_ROTATE_FALLBACK)
@@ -912,7 +925,14 @@ def _status_403(c: _Ctx) -> Verdict:
     # OpenRouter 403 "key limit exceeded" and similar plan/credit exhaustion are billing.
     xai_spend = c.provider_slug == "xai-oauth" and c.code == _XAI_SPENDING_LIMIT_ERROR_CODE
     billing = xai_spend or any(p in c.msg for p in ("key limit exceeded", "spending limit") + _BILLING_PATTERNS)
-    return _V_BILLING if billing else _V_AUTH_FALLBACK
+    if billing:
+        return _V_BILLING
+    # A WAF/CDN in front of the provider answered, not the provider: the credential never
+    # reached it, so key guidance and credential rotation are wrong (#53099, #70566). Gated on
+    # 403 and on established block/challenge markers; any other 403 stays auth.
+    if any(p in c.msg for p in _UPSTREAM_BLOCKED_PATTERNS):
+        return _V_UPSTREAM_BLOCKED
+    return _V_AUTH_FALLBACK
 
 
 def _status_404(c: _Ctx) -> Verdict:
diff --git a/agent/turn_failure_copy.py b/agent/turn_failure_copy.py
index 7369458406..272888fd8d 100644
--- a/agent/turn_failure_copy.py
+++ b/agent/turn_failure_copy.py
@@ -168,6 +168,11 @@ _NONRETRYABLE_COPY: Dict[str, str] = {
         "{label}'s account settings don't allow this model for your request, so it didn't "
         "answer. Check the provider's data/privacy settings, or switch models with /model."
     ),
+    FailoverReason.upstream_blocked.value: (
+        "A firewall/CDN in front of {label} blocked the request before it reached the model, so "
+        "your key is probably fine. Set a custom User-Agent via the provider's extra_headers, check "
+        "the proxy/WAF rules, or switch providers with /model."
+    ),
 }
 _NONRETRYABLE_DEFAULT_COPY = (
     "{label} rejected the request and retrying won't help. Pick another model with /model, "
@@ -202,6 +207,7 @@ FAILURE_CAUSE_GLOSS: Dict[str, str] = {
     "billing_unverified": "the AI model service says the account's usage or credit limit is reached",
     FailoverReason.auth.value: "the AI model service rejected the sign-in",
     FailoverReason.auth_permanent.value: "the AI model service rejected the sign-in",
+    FailoverReason.upstream_blocked.value: "a firewall/CDN in front of the AI model service blocked the request",
     FailoverReason.model_not_found.value: "the model {subject} uses was not found at the AI model service",
     FailoverReason.content_policy_blocked.value: "the AI model service's safety filter rejected the request",
     "context_overflow": "{possessive} request grew too large for the model",
diff --git a/agent/turn_recovery.py b/agent/turn_recovery.py
index 1c338b12b5..731e54bf4e 100644
--- a/agent/turn_recovery.py
+++ b/agent/turn_recovery.py
@@ -827,6 +827,7 @@ def _welcome_outage_copy(base_url: Any, classified: Any, *, anonymous: bool = Fa
 # Terminal status label per non-retryable reason (default names the HTTP status).
 _NONRETRYABLE_LABELS = {
     FailoverReason.content_policy_blocked: "The provider's safety filter refused this request",
+    FailoverReason.upstream_blocked: "A firewall/CDN in front of the provider blocked this request",
     FailoverReason.ssl_cert_verification: "The provider's security certificate could not be verified",
     # Only reached after the one-shot image shrink ran (recover_after_classification sets the flag first).
     FailoverReason.image_too_large: "Request still exceeded the provider's size limit after shrinking images",
@@ -889,6 +890,16 @@ def nonretryable_client_error_result(
             _vlines(agent, f"      Did you mean '{_prefix_suggestion}'? It looks like the vendor prefix is missing.")
     elif classified.reason not in _NONRETRYABLE_LABELS:
         _vlines(agent, f"   💡 Fix: pick another model (/model), or check `{display_hermes_home()}/logs/agent.log`.")
+    # A WAF/CDN block (#53099, #70566): the key never reached the provider; the usual cause
+    # is the SDK User-Agent, which the per-provider extra_headers override.
+    if classified.reason == FailoverReason.upstream_blocked:
+        _vlines(
+            agent,
+            "   💡 The endpoint's firewall/CDN blocked the request before it reached the model — your key",
+            "      and model access are probably fine. Relays often reject the SDK's default User-Agent:",
+            "      set `extra_headers: {User-Agent: HermesAgent/1.0}` on the custom_providers entry",
+            "      (or `model.default_headers`), or check the proxy/WAF rules and your network.",
+        )
     # Content-policy blocks: the provider refused this prompt, so recovery is a rephrase
     # or another model, not key/retry advice.
     if classified.reason == FailoverReason.content_policy_blocked:
diff --git a/hermes_cli/cli_chat_error_copy.py b/hermes_cli/cli_chat_error_copy.py
index 2327b21616..81ce86c460 100644
--- a/hermes_cli/cli_chat_error_copy.py
+++ b/hermes_cli/cli_chat_error_copy.py
@@ -17,6 +17,7 @@ _REASON_COPY: dict[str, str] = {
     "model_not_found": "'{model}' isn't available on {provider}. Run /model to pick a valid model.",
     "rate_limit": "Rate limited by {provider}; wait a minute or /model to switch.",
     "upstream_rate_limit": "Rate limited by {provider}; wait a minute or /model to switch.",
+    "upstream_blocked": "A firewall/CDN in front of {provider} blocked the request (not your key). Set a User-Agent via extra_headers, or /model to switch.",
     "overloaded": "{provider} is overloaded right now. Send /retry in a moment, or /model to switch.",
     "server_error": "{provider} had an internal error. Send /retry in a moment, or /model to switch.",
     "timeout": "{provider} did not answer in time. Send /retry, or /model to switch.",
diff --git a/tests/agent/test_anthropic_adapter_extra_headers.py b/tests/agent/test_anthropic_adapter_extra_headers.py
new file mode 100644
index 0000000000..bf1aa738dd
--- /dev/null
+++ b/tests/agent/test_anthropic_adapter_extra_headers.py
@@ -0,0 +1,32 @@
+"""Anthropic-messages clients honour ``custom_providers[].extra_headers`` (#24293, #9721).
+
+The OpenAI-wire clients apply the per-provider headers; ``build_anthropic_client`` used to skip
+them, so a relay behind a WAF that rejects the SDK User-Agent kept 403ing in anthropic_messages mode.
+"""
+from unittest.mock import patch
+
+from agent.anthropic_adapter import build_anthropic_client
+
+_ROUTE = "https://proxy.example.com/v1"
+_CONFIG = {"custom_providers": [{
+    "name": "wafproxy", "base_url": _ROUTE, "api_mode": "anthropic_messages",
+    "extra_headers": {"User-Agent": "HermesAgent/1.0", "X-Privacy-Tier": "enterprise"},
+}]}
+
+
+def _build(route):
+    with patch("agent.anthropic_adapter._require_sdk") as sdk, patch("hermes_cli.config.load_config", return_value=_CONFIG):
+        build_anthropic_client("sk-test", route)
+    return sdk.return_value.Anthropic.call_args.kwargs["default_headers"]
+
+
+def test_matching_route_merges_extra_headers_after_betas():
+    headers = _build(_ROUTE)
+    assert headers["User-Agent"] == "HermesAgent/1.0"
+    assert headers["X-Privacy-Tier"] == "enterprise"
+    assert "anthropic-beta" in headers  # provider headers add to, not replace, the beta set
+
+
+def test_other_route_does_not_inherit_extra_headers():
+    headers = _build("https://other.example.com/v1")
+    assert "User-Agent" not in headers and "X-Privacy-Tier" not in headers
diff --git a/tests/agent/test_error_classifier.py b/tests/agent/test_error_classifier.py
index c5a5e21fe5..059d633b57 100644
--- a/tests/agent/test_error_classifier.py
+++ b/tests/agent/test_error_classifier.py
@@ -60,7 +60,7 @@ class TestFailoverReason:
     def test_enum_members_exist(self):
         expected = {
             "auth", "auth_permanent", "billing", "rate_limit",
-            "upstream_rate_limit",
+            "upstream_rate_limit", "upstream_blocked",
             "overloaded", "server_error", "timeout",
             "ssl_cert_verification",
             "context_overflow", "payload_too_large", "image_too_large",
diff --git a/tests/agent/test_error_classifier_upstream_blocked.py b/tests/agent/test_error_classifier_upstream_blocked.py
new file mode 100644
index 0000000000..faace46d05
--- /dev/null
+++ b/tests/agent/test_error_classifier_upstream_blocked.py
@@ -0,0 +1,34 @@
+"""A 403 written by a WAF/CDN in front of the provider is not an API-key rejection (#53099, #70566).
+
+A relay that blocks the SDK User-Agent answers ``403 Your request was blocked.``; Cloudflare's
+browser challenge answers 403 HTML. Both used to classify as ``auth`` and print key guidance.
+"""
+import pytest
+
+from agent.error_classifier import FailoverReason, classify_api_error
+
+
+class _APIError(Exception):
+    def __init__(self, message, status_code):
+        super().__init__(message)
+        self.status_code = status_code
+
+
+@pytest.mark.parametrize("body", [
+    "Error code: 403 - Your request was blocked.",
+    "<!doctype html><html><body>Enable JavaScript and cookies to continue</body></html>",
+    "<!doctype html><html><script src='/cdn-cgi/challenge-platform/h/g/orchestrate/chl_page'></script></html>",
+])
+def test_403_waf_block_is_upstream_blocked_not_auth(body):
+    result = classify_api_error(_APIError(body, 403), provider="openai-api")
+    assert result.reason == FailoverReason.upstream_blocked
+    assert result.retryable is False and result.should_fallback is True
+    assert result.should_rotate_credential is False and result.is_auth is False
+
+
+@pytest.mark.parametrize("body, status, reason", [
+    ("<html><title>Forbidden</title><body>Access denied</body></html>", 403, FailoverReason.auth),
+    ("<html>Enable JavaScript and cookies to continue</html>", 401, FailoverReason.auth),
+])
+def test_generic_403_and_all_401_keep_auth(body, status, reason):
+    assert classify_api_error(_APIError(body, status), provider="openai-api").reason == reason
diff --git a/tui_gateway/user_messages.py b/tui_gateway/user_messages.py
index 61afbb8eee..7822c05874 100644
--- a/tui_gateway/user_messages.py
+++ b/tui_gateway/user_messages.py
@@ -21,6 +21,7 @@ _TURN_ERROR_CODE_COPY: dict[str, tuple[str, str]] = {
     "billing_unverified": ("The model provider reports no credit left", "Top up the account or switch with /model."),
     "rate_limit": ("The model provider is rate-limiting requests", "Wait a moment, then /retry."),
     "upstream_rate_limit": ("The model provider is rate-limiting requests", "Wait a moment, then /retry."),
+    "upstream_blocked": ("A firewall/CDN in front of the model provider blocked the request", "Set a User-Agent via the provider's extra_headers, or switch with /model."),
     "overloaded": ("The model provider is overloaded", "Wait a moment, then /retry."),
     "server_error": ("The model provider had an internal error", "Wait a moment, then /retry."),
     "timeout": ("The model provider did not answer in time", "Try /retry; if it keeps happening, switch with /model."),
diff --git a/ui-tui/src/app/userMessages.ts b/ui-tui/src/app/userMessages.ts
index 881620833f..76da58f52c 100644
--- a/ui-tui/src/app/userMessages.ts
+++ b/ui-tui/src/app/userMessages.ts
@@ -241,6 +241,10 @@ const TURN_CODE_COPY: Record<string, [string, string]> = {
     "Check the endpoint's certificate, then /retry."
   ],
   timeout: ['The model provider did not answer in time', 'Try /retry; if it keeps happening, switch with /model.'],
+  upstream_blocked: [
+    'A firewall/CDN in front of the model provider blocked the request',
+    "Set a User-Agent via the provider's extra_headers, or switch with /model."
+  ],
   upstream_rate_limit: ['The model provider is rate-limiting requests', 'Wait a moment, then /retry.']
 }
 
diff --git a/website/docs/user-guide/configuring-models.md b/website/docs/user-guide/configuring-models.md
index 94345b2d61..e830d871b8 100644
--- a/website/docs/user-guide/configuring-models.md
+++ b/website/docs/user-guide/configuring-models.md
@@ -192,7 +192,7 @@ providers:
       CF-Access-Client-Secret: "yyyy"
 ```
 
-Header values routinely carry credentials — Hermes never logs them. `extra_headers` applies to OpenAI-compatible routes; the `anthropic_messages` and `bedrock_converse` API modes do not use it.
+Header values routinely carry credentials — Hermes never logs them. `extra_headers` applies to OpenAI-compatible routes and to `anthropic_messages` routes (the main client, `/model` switches, rebuilds and auxiliary clients alike); `bedrock_converse` does not use it. A relay behind a WAF that rejects the SDK's default `User-Agent` (403 "Your request was blocked" or a browser-challenge page) is the typical reason to set one — Hermes reports such a 403 as a firewall/CDN block rather than an API-key rejection.
 
 **`discover_models`** — set to `false` (default `true`) to skip querying the endpoint's `/models` listing and use only the `models` you configured on the entry. Handy for gateways whose model listing is slow, unreliable, or noisy: