Merge pull request #39964 from BerriAI/litellm_lit_7050_redact_failure_traceback

fix(proxy): redact provider keys from pass-through failure tracebacks

litellm/litellm_core_utils/litellm_logging.py

@@ -5675,8 +5675,8 @@ class StandardLoggingPayloadSetup:
             error_code=error_status,
             error_class=error_class,
             llm_provider=_llm_provider_in_exception,
-            traceback=traceback_info,
-            error_message=error_message,
+            traceback=_redact_string(traceback_info),
+            error_message=_redact_string(error_message),
             error_rate_limit_category=rate_limit_category,
             error_rate_limit_type=rate_limit_type,
             error_budget_entity_type=budget_error.entity_type if budget_error else None,

litellm/proxy/utils.py

@@ -2568,6 +2568,8 @@ class ProxyLogging:
         # Remove before callbacks iterate — not serialisable
         request_data.pop("litellm_logging_obj", None)
 
+        redacted_traceback_str: Final = _redact_string(traceback_str) if traceback_str is not None else None
+
         # Track the first HTTPException returned or raised by any callback
         transformed_exception: HTTPException | None = None
 
@@ -2586,7 +2588,7 @@ class ProxyLogging:
                             request_data=request_data,
                             user_api_key_dict=user_api_key_dict,
                             original_exception=original_exception,
-                            traceback_str=traceback_str,
+                            traceback_str=redacted_traceback_str,
                         )
                         # If callback returned an HTTPException, use it (first one wins)
                         if isinstance(hook_result, HTTPException) and transformed_exception is None:

tests/test_litellm/litellm_core_utils/test_litellm_logging.py

@@ -6457,3 +6457,32 @@ def test_get_status_fields_ranks_guardrail_flagged_between_success_and_intervene
     assert _get_status_fields(
         "success", [flagged, {"guardrail_status": "guardrail_intervened"}], None
     )["guardrail_status"] == "guardrail_intervened"
+
+
+def test_get_error_information_redacts_provider_key_from_upstream_url():
+    """A pass-through upstream failure logs the httpx traceback, whose message
+    quotes the upstream URL with the provider key in its query string. That
+    key must never reach spend logs or logging callbacks."""
+    import traceback
+
+    from litellm.litellm_core_utils.litellm_logging import StandardLoggingPayloadSetup
+
+    provider_key = "AIza" + "S" * 35
+    upstream_url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini:generateContent?key={provider_key}"
+    response = httpx.Response(400, request=httpx.Request("POST", upstream_url))
+    try:
+        response.raise_for_status()
+    except httpx.HTTPStatusError as caught:
+        upstream_error = caught
+        upstream_traceback = traceback.format_exc()
+    assert provider_key in upstream_traceback
+
+    result = StandardLoggingPayloadSetup.get_error_information(
+        original_exception=upstream_error, traceback_str=upstream_traceback
+    )
+
+    assert provider_key not in result["traceback"]
+    assert provider_key not in result["error_message"]
+    assert "REDACTED" in result["traceback"]
+    assert "REDACTED" in result["error_message"]
+    assert result["error_code"] == "400"

tests/test_litellm/proxy/test_proxy_utils.py

@@ -6,7 +6,8 @@ from fastapi import HTTPException
 
 from litellm.caching.caching import DualCache
 from litellm.integrations.custom_guardrail import CustomGuardrail
-from litellm.proxy._types import ProxyErrorTypes
+from litellm.integrations.custom_logger import CustomLogger
+from litellm.proxy._types import ProxyErrorTypes, UserAPIKeyAuth
 from litellm.proxy.utils import ProxyLogging
 from litellm.types.guardrails import GuardrailEventHooks
 
@@ -1919,3 +1920,55 @@ async def test_proxy_only_error_5xx_keeps_traceback_and_runs_sync_callbacks(monk
         Logging.failure_handler = orig_sync_failure
 
     assert "test_proxy_utils" in captured["async_traceback"]
+
+
+class _TracebackRecordingLogger(CustomLogger):
+    def __init__(self) -> None:
+        super().__init__()
+        self.received_traceback: str | None = None
+
+    async def async_post_call_failure_hook(
+        self,
+        request_data: dict,
+        original_exception: Exception,
+        user_api_key_dict: UserAPIKeyAuth,
+        traceback_str: str | None = None,
+    ) -> HTTPException | None:
+        self.received_traceback = traceback_str
+        return None
+
+
+@pytest.mark.asyncio
+async def test_post_call_failure_hook_redacts_traceback_before_callbacks(monkeypatch):
+    """A pass-through upstream failure hands the hook the httpx traceback, whose
+    message quotes the upstream URL with the provider key in its query string.
+    Every callback, custom loggers included, must receive it redacted."""
+    import traceback
+    from unittest.mock import AsyncMock, patch
+
+    import httpx
+
+    provider_key = "AIza" + "S" * 35
+    upstream_url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini:generateContent?key={provider_key}"
+    response = httpx.Response(400, request=httpx.Request("POST", upstream_url))
+    try:
+        response.raise_for_status()
+    except httpx.HTTPStatusError:
+        upstream_traceback = traceback.format_exc()
+    assert provider_key in upstream_traceback
+
+    recorder = _TracebackRecordingLogger()
+    monkeypatch.setattr(litellm, "callbacks", [recorder])
+    proxy_logging_obj = ProxyLogging(user_api_key_cache=DualCache())
+    proxy_logging_obj.alert_types = []
+    with patch.object(proxy_logging_obj, "update_request_status", new=AsyncMock()):
+        await proxy_logging_obj.post_call_failure_hook(
+            request_data={"metadata": {}},
+            original_exception=HTTPException(status_code=400, detail="Upstream passthrough request failed with status 400"),
+            user_api_key_dict=UserAPIKeyAuth(),
+            traceback_str=upstream_traceback,
+        )
+
+    assert recorder.received_traceback is not None
+    assert provider_key not in recorder.received_traceback
+    assert "REDACTED" in recorder.received_traceback