fix(batches): mask pre-signed request auth headers before raw-request logging

A pre-signed batch/file request (Mistral, Bedrock) carries its auth header
inside the transformed request body, which pre_call logs verbatim into
raw_request_typed_dict and raw-request callbacks, leaking the provider key.
Mask the nested headers channel before handing the request to pre_call.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

litellm/llms/custom_httpx/llm_http_handler.py

@@ -278,6 +278,23 @@ def _has_pre_call_deployment_hook(logging_obj: LiteLLMLoggingObj) -> bool:
     return False
 
 
+def _mask_presigned_request_headers(transformed_request: bytes | str | dict) -> bytes | str | dict:
+    """A pre-signed request carries its auth inside its own ``headers`` key, which
+    logging treats as request body (only the top-level headers channel gets masked),
+    so mask it here before the request is handed to ``pre_call``."""
+    if not isinstance(transformed_request, dict):
+        return transformed_request
+    request_headers: Final = transformed_request.get("headers")
+    if not isinstance(request_headers, dict):
+        return transformed_request
+
+    from litellm.litellm_core_utils.litellm_logging import (
+        _get_masked_values,  # pyright: ignore[reportPrivateUsage]  # the shared header-masking helper has no public name
+    )
+
+    return {**transformed_request, "headers": _get_masked_values(request_headers)}
+
+
 def _aws_signing_overrides(optional_params: Mapping[str, Any], litellm_params: Mapping[str, Any]) -> Mapping[str, Any]:
     return MappingProxyType(
         {
@@ -3692,7 +3709,7 @@ class BaseLLMHTTPHandler:
                 "complete_input_dict": (
                     "<streaming media upload>"
                     if isinstance(transformed_request, dict) and "streaming_media_upload" in transformed_request
-                    else transformed_request
+                    else _mask_presigned_request_headers(transformed_request)
                 ),
                 "api_base": api_base,
                 "headers": headers,
@@ -4115,7 +4132,7 @@ class BaseLLMHTTPHandler:
             input="",
             api_key="",
             additional_args={
-                "complete_input_dict": transformed_request,
+                "complete_input_dict": _mask_presigned_request_headers(transformed_request),
                 "api_base": api_base,
                 "headers": headers,
             },
@@ -4194,7 +4211,7 @@ class BaseLLMHTTPHandler:
             input="",
             api_key="",
             additional_args={
-                "complete_input_dict": transformed_request,
+                "complete_input_dict": _mask_presigned_request_headers(transformed_request),
                 "api_base": api_base,
                 "headers": headers,
                 "batch_id": batch_id,

tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py

@@ -2761,6 +2761,68 @@ def test_direct_vector_store_search_debug_log_omits_stored_credentials(caplog, i
     assert "sk-embedding-s3cret" not in logged
 
 
+@pytest.mark.asyncio
+async def test_async_retrieve_batch_masks_presigned_auth_header_in_raw_request_log():
+    """Regression: a pre-signed retrieve-batch request (Mistral, Bedrock) embeds its auth
+    header inside the transformed request, which pre_call logs verbatim as the raw request
+    body, so the provider key landed unmasked in raw_request_typed_dict and every
+    raw-request callback."""
+    from litellm.litellm_core_utils.litellm_logging import Logging as LitellmLogging
+    from litellm.llms.mistral.batches.transformation import MistralBatchesConfig
+
+    provider_key = "mistral-s3cret-provider-key-123456"
+    job_payload = {
+        "id": "batch-1",
+        "input_files": ["file-1"],
+        "endpoint": "/v1/ocr",
+        "model": "mistral-ocr-latest",
+        "status": "SUCCESS",
+        "created_at": 1_757_400_000,
+    }
+    sent_requests = []
+
+    def _capture(request: httpx.Request) -> httpx.Response:
+        sent_requests.append(request)
+        return httpx.Response(200, json=job_payload)
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(_capture))
+
+    logging_obj = LitellmLogging(
+        model="mistral/mistral-ocr-latest",
+        messages=[],
+        stream=False,
+        call_type="batch_retrieve",
+        start_time=time.time(),
+        litellm_call_id="batch-retrieve-call-id",
+        function_id="batch-retrieve-function-id",
+        log_raw_request_response=True,
+    )
+    logging_obj.update_environment_variables(
+        model="mistral/mistral-ocr-latest",
+        optional_params={},
+        litellm_params={"litellm_call_id": "batch-retrieve-call-id", "metadata": {}},
+    )
+
+    result = await BaseLLMHTTPHandler().retrieve_batch(
+        batch_id="batch-1",
+        litellm_params={"api_key": provider_key},
+        provider_config=MistralBatchesConfig(),
+        headers={},
+        api_base=None,
+        api_key=provider_key,
+        logging_obj=logging_obj,
+        _is_async=True,
+        client=client,
+        model="mistral/mistral-ocr-latest",
+    )
+
+    assert result.id == "batch-1"
+    assert sent_requests[0].headers["Authorization"] == f"Bearer {provider_key}"
+    raw_request_body = logging_obj.model_call_details["raw_request_typed_dict"]["raw_request_body"]
+    assert provider_key not in json.dumps(raw_request_body)
+
+
 @pytest.mark.asyncio
 async def test_async_anthropic_messages_handler_carries_deployment_vertex_location_for_pricing(monkeypatch):
     """