From 3fa2bd34bb0a044a1b0dfb45913f85cf0837f78f Mon Sep 17 00:00:00 2001
From: Varshith <kvarshithgowda@gmail.com>
Date: Fri, 22 May 2026 09:52:03 -0500
Subject: [PATCH] fix(fallbacks): ignore preserve_litellm_internal_headers for
 raw httpx.Headers inputs

---
 litellm/litellm_core_utils/core_helpers.py    | 24 +++++++++++++------
 .../litellm_core_utils/test_fallback_utils.py | 20 ++++++++++++++++
 2 files changed, 37 insertions(+), 7 deletions(-)

diff --git a/litellm/litellm_core_utils/core_helpers.py b/litellm/litellm_core_utils/core_helpers.py
index 536ac309f42..8b96d8355de 100644
--- a/litellm/litellm_core_utils/core_helpers.py
+++ b/litellm/litellm_core_utils/core_helpers.py
@@ -247,15 +247,25 @@ def process_response_headers(
     preserve_litellm_internal_headers: bool = False,
 ) -> dict:
     """
-    `preserve_litellm_internal_headers` must only be True when the input dict is
-    a LiteLLM-owned structure (e.g. `_hidden_params["additional_headers"]` that
-    has already been through one round of processing). For raw upstream
-    provider headers it must remain False, otherwise a malicious provider
-    returning `x-litellm-*` could spoof LiteLLM-internal markers
-    (e.g. `x-litellm-attempted-fallbacks`).
+    `preserve_litellm_internal_headers` must only be True when the input is a
+    LiteLLM-owned dict (e.g. `_hidden_params["additional_headers"]` that has
+    already been through one round of processing).  For raw upstream provider
+    headers — whether passed as `httpx.Headers` or a plain dict — it must
+    remain False, otherwise a malicious provider returning `x-litellm-*` could
+    spoof LiteLLM-internal markers (e.g. `x-litellm-attempted-fallbacks`).
+
+    When the input is an `httpx.Headers` object the flag is always treated as
+    False regardless of what the caller requested, because `httpx.Headers` is
+    always a raw provider response and can never be LiteLLM-owned.
     """
     from litellm.types.utils import OPENAI_RESPONSE_HEADERS
 
+    # Raw httpx.Headers objects come directly from provider HTTP responses and
+    # must never be treated as LiteLLM-owned, regardless of caller intent.
+    _preserve = preserve_litellm_internal_headers and isinstance(
+        response_headers, dict
+    )
+
     openai_headers = {}
     processed_headers = {}
     additional_headers = {}
@@ -267,7 +277,7 @@ def process_response_headers(
             "llm_provider-"
         ):  # return raw provider headers (incl. openai-compatible ones)
             processed_headers[k] = v
-        elif preserve_litellm_internal_headers and k.startswith("x-litellm-"):
+        elif _preserve and k.startswith("x-litellm-"):
             # LiteLLM's own internal headers (e.g. x-litellm-attempted-fallbacks,
             # x-litellm-model-group) are not LLM provider headers and must not be
             # prefixed. Downstream consumers (proxy override, callers checking
diff --git a/tests/test_litellm/litellm_core_utils/test_fallback_utils.py b/tests/test_litellm/litellm_core_utils/test_fallback_utils.py
index 142602ccb27..90a61696e9d 100644
--- a/tests/test_litellm/litellm_core_utils/test_fallback_utils.py
+++ b/tests/test_litellm/litellm_core_utils/test_fallback_utils.py
@@ -1,6 +1,7 @@
 """Tests for litellm.litellm_core_utils.fallback_utils."""
 
 import pytest
+import httpx
 
 import litellm
 from litellm.litellm_core_utils.core_helpers import process_response_headers
@@ -147,3 +148,22 @@ def test_process_response_headers_prefixes_x_litellm_from_raw_provider():
     assert "x-litellm-attempted-fallbacks" not in result
     assert result["llm_provider-x-litellm-attempted-fallbacks"] == 99
     assert result["llm_provider-x-stainless-arch"] == "arm64"
+
+
+def test_process_response_headers_ignores_preserve_flag_for_httpx_headers():
+    """
+    Some providers store raw httpx.Headers directly in _hidden_params["additional_headers"]
+    without a prior normalization pass. If preserve_litellm_internal_headers=True were
+    honored for httpx.Headers inputs, a provider returning x-litellm-attempted-fallbacks
+    could spoof it as a bare LiteLLM-internal marker and make the proxy skip
+    stamping the correct response model. The flag must be ignored for httpx.Headers.
+    """
+    raw = httpx.Headers(
+        {
+            "x-litellm-attempted-fallbacks": "1",
+            "content-type": "application/json",
+        }
+    )
+    result = process_response_headers(raw, preserve_litellm_internal_headers=True)
+    assert "x-litellm-attempted-fallbacks" not in result
+    assert result["llm_provider-x-litellm-attempted-fallbacks"] == "1"