Merge 85e61375d9 into 2dccc0dc79

litellm/proxy/litellm_pre_call_utils.py

@@ -313,6 +313,10 @@ _UNTRUSTED_ROOT_CONTROL_FIELDS: Final = (
 )
 
 _UNTRUSTED_METADATA_CONTROL_FIELDS: Final = (
+    # Proxy-authoritative clientside-credential opt-in scope; the proxy
+    # overwrites it in add_litellm_data_to_request, so a caller-forged value
+    # must never survive to the router.
+    "litellm_proxy_clientside_credential_scope",
     "disable_global_guardrails",
     "disable_global_guardrail",
     "opted_out_global_guardrails",
@@ -1970,7 +1974,15 @@ async def add_litellm_data_to_request(
 
     """
 
-    from litellm.proxy.proxy_server import llm_router, premium_user
+    # The general_settings alias is the fallback for callers that don't pass
+    # it: needed to stamp the clientside-credential opt-in scope below.
+    from litellm.proxy.proxy_server import (
+        general_settings as _proxy_general_settings,
+    )
+    from litellm.proxy.proxy_server import (
+        llm_router,
+        premium_user,
+    )
     from litellm.types.proxy.litellm_pre_call_utils import RedactedDict, SecretFields
 
     # Strip internal-only keys from user input before the proxy sets its own.
@@ -2255,6 +2267,28 @@ async def add_litellm_data_to_request(
     )
     data[_metadata_variable_name]["litellm_api_version"] = version
 
+    # Stamp the authoritative clientside-credential opt-in scope for this
+    # request: "proxy_wide" when general_settings.allow_client_side_credentials
+    # is true, else "per_model". The router re-validates per-deployment opt-in
+    # whenever these kwargs are re-dispatched (server-side fallbacks) — without
+    # this, a fallback target that never opted in would get its api_base
+    # overridden to the caller's URL while keeping its own api_key.
+    # Any caller-supplied value for this key was stripped above
+    # (_UNTRUSTED_METADATA_CONTROL_FIELDS), so this value cannot be forged.
+    from litellm.router_utils.clientside_credential_handler import (
+        PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_METADATA_KEY,
+        PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PER_MODEL,
+        PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PROXY_WIDE,
+    )
+
+    _clientside_scope_settings: Final = general_settings if general_settings is not None else _proxy_general_settings
+    data[_metadata_variable_name][PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_METADATA_KEY] = (
+        PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PROXY_WIDE
+        if isinstance(_clientside_scope_settings, dict)
+        and _clientside_scope_settings.get("allow_client_side_credentials") is True
+        else PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PER_MODEL
+    )
+
     if general_settings is not None:
         data[_metadata_variable_name]["global_max_parallel_requests"] = general_settings.get(
             "global_max_parallel_requests", None

litellm/router.py

@@ -172,6 +172,7 @@ from litellm.router_utils.client_initalization_utils import InitalizeCachedClien
 from litellm.router_utils.clientside_credential_handler import (
     get_dynamic_litellm_params,
     is_clientside_credential,
+    strip_clientside_credentials_without_deployment_opt_in,
 )
 from litellm.router_utils.common_utils import (
     _is_proxy_admin_request,
@@ -3924,6 +3925,12 @@ class Router:
         deployment_litellm_model_name = deployment["litellm_params"]["model"]
         deployment_api_base = deployment["litellm_params"].get("api_base")
         deployment_model_name: Final = deployment["model_name"]
+        # Re-dispatches (server-side fallbacks) reuse these kwargs against a
+        # deployment that never saw the auth-time opt-in check. Drop any
+        # clientside credential key this deployment did not opt in to before
+        # the clientside-credential branch can forward them (which would
+        # override the deployment's api_base while keeping its api_key).
+        strip_clientside_credentials_without_deployment_opt_in(deployment=deployment, kwargs=kwargs)
         if is_clientside_credential(request_kwargs=kwargs):
             deployment_pydantic_obj: Final = self._handle_clientside_credential(
                 deployment=deployment, kwargs=kwargs, function_name=function_name

litellm/router_utils/clientside_credential_handler.py

@@ -11,10 +11,24 @@ If given, generate a unique model_id for the deployment.
 Ensures cooldowns are applied correctly.
 """
 
+import re
+from collections.abc import Mapping
 from typing import Final
 
 clientside_credential_keys: Final = ["api_key", "api_base", "base_url"]
 
+# Metadata key the proxy stamps with the admin opt-in scope that authorized
+# the caller's clientside credentials at auth time. Consumed by
+# strip_clientside_credentials_without_deployment_opt_in below to re-validate
+# per-deployment opt-in when a re-dispatch (server-side router fallback) would
+# otherwise forward those credentials to a deployment that never opted in.
+# User-supplied values for this key are stripped by the proxy (see
+# litellm/proxy/litellm_pre_call_utils.py _UNTRUSTED_METADATA_CONTROL_FIELDS),
+# so the scope cannot be forged by a caller.
+PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_METADATA_KEY: Final = "litellm_proxy_clientside_credential_scope"
+PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PROXY_WIDE: Final = "proxy_wide"
+PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PER_MODEL: Final = "per_model"
+
 
 def _admin_config_fields_to_clear_on_base_override() -> list[str]:
     """
@@ -103,3 +117,99 @@ def get_dynamic_litellm_params(litellm_params: dict, request_kwargs: dict) -> di
                 litellm_params[field] = request_kwargs[field]
 
     return litellm_params
+
+
+def _clientside_param_allowed_for_deployment(
+    param: str,
+    request_body_value: object,
+    configurable_clientside_auth_params: object,
+) -> bool:
+    """
+    Mirror of litellm.proxy.auth.auth_utils._is_param_allowed (kept here
+    because router_utils cannot import from the proxy package without a
+    circular import). A param is allowed when the deployment's
+    ``configurable_clientside_auth_params`` names it, or — for ``api_base``
+    only — a ``{"api_base": <pattern>}`` dict entry regex/equal-matches the
+    caller-supplied value.
+    """
+    if configurable_clientside_auth_params is None:
+        return False
+
+    for item in configurable_clientside_auth_params:
+        if isinstance(item, str) and param == item:
+            return True
+        if isinstance(item, dict) and param == "api_base" and isinstance(request_body_value, str):
+            pattern = item.get("api_base")
+            if isinstance(pattern, str) and (re.match(pattern, request_body_value) or pattern == request_body_value):
+                return True
+
+    return False
+
+
+def strip_clientside_credentials_without_deployment_opt_in(deployment: Mapping[str, object], kwargs: dict) -> None:
+    """
+    Re-validate the proxy's per-deployment clientside-credential opt-in at
+    (re-)dispatch time.
+
+    Auth time (litellm.proxy.auth.auth_utils._check_banned_params) validates
+    caller-supplied clientside credentials (api_key / api_base / base_url)
+    against the model the caller DECLARED only. Server-side router fallbacks
+    re-dispatch the same kwargs to a DIFFERENT deployment, which must re-consent:
+    without this check, a deployment that never opted in gets its api_base
+    overridden to the caller's URL while keeping its own api_key, exfiltrating
+    that deployment's provider key to the caller-chosen host.
+
+    Scoping (deliberately narrow so the SDK router's per-call credential
+    feature is unchanged):
+    - No scope stamp in kwargs metadata (plain SDK completion call, or a proxy
+      route that never stamped one): unchanged behavior.
+    - ``proxy_wide`` (general_settings.allow_client_side_credentials = true):
+      the admin opted every deployment in; unchanged behavior.
+    - ``per_model``: each clientside credential key in kwargs must be opted in
+      by the deployment being dispatched (its model_info /
+      litellm_params ``configurable_clientside_auth_params``); keys that are
+      not are stripped so the dispatch uses the deployment's own config.
+
+    Metadata buckets: the proxy stamps the scope into exactly one of
+    ``metadata`` / ``litellm_metadata`` (``_get_metadata_variable_name``).
+    On LITELLM_METADATA_ROUTES (/v1/messages, responses, batches, bedrock,
+    files) and the thread/assistant routes that bucket is
+    ``litellm_metadata``, while a caller-supplied provider-facing ``metadata``
+    object survives as a SECOND kwargs bucket. This helper therefore inspects
+    every bucket — reading only one let an unstamped caller ``metadata``
+    object shadow the stamp and skip stripping on fallback re-dispatch.
+    Non-dict buckets (e.g. a JSON-encoded string) cannot carry the stamp and
+    are skipped. Caller-forged scope values are stripped upstream
+    (_UNTRUSTED_METADATA_CONTROL_FIELDS), so any value found is
+    proxy-authored; if buckets ever disagree, fail closed to the most
+    restrictive scope (``per_model``), mirroring the both-bucket loops the
+    proxy uses at its other security boundaries.
+    """
+    scope: str | None = None
+    for metadata_key in ("metadata", "litellm_metadata"):
+        metadata_bucket: object = kwargs.get(metadata_key)
+        if not isinstance(metadata_bucket, dict):
+            continue
+        bucket_scope: object = metadata_bucket.get(PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_METADATA_KEY)
+        if bucket_scope == PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PER_MODEL:
+            scope = PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PER_MODEL
+            break
+        if bucket_scope == PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PROXY_WIDE and scope is None:
+            scope = PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PROXY_WIDE
+    if scope != PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_PER_MODEL:
+        return
+
+    allowed_params: object = None
+    model_info: object = deployment.get("model_info")
+    if isinstance(model_info, Mapping):
+        allowed_params = model_info.get("configurable_clientside_auth_params")
+    if allowed_params is None:
+        litellm_params: object = deployment.get("litellm_params")
+        if isinstance(litellm_params, Mapping):
+            allowed_params = litellm_params.get("configurable_clientside_auth_params")
+        else:
+            allowed_params = getattr(litellm_params, "configurable_clientside_auth_params", None)
+
+    for key in clientside_credential_keys:
+        if key in kwargs and not _clientside_param_allowed_for_deployment(key, kwargs[key], allowed_params):
+            kwargs.pop(key, None)

tests/test_litellm/router_utils/test_clientside_credential_handler.py

@@ -0,0 +1,148 @@
+"""
+Regression tests for the clientside-credential fallback re-validation scope
+lookup (see litellm/router_utils/clientside_credential_handler.py).
+
+The proxy stamps the opt-in scope into exactly one metadata bucket
+(``_get_metadata_variable_name``): ``litellm_metadata`` on
+LITELLM_METADATA_ROUTES (/v1/messages, responses, batches, bedrock, files)
+and the thread/assistant routes, ``metadata`` everywhere else — while a
+caller-supplied provider-facing ``metadata`` object can survive as a second
+kwargs bucket on those routes. The strip helper must consult EVERY bucket so
+an unstamped caller bucket cannot shadow the proxy stamp (PR 40001 review:
+"High: Duplicate metadata bypasses fallback credential stripping").
+"""
+
+from litellm.router_utils.clientside_credential_handler import (
+    PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_METADATA_KEY,
+    strip_clientside_credentials_without_deployment_opt_in,
+)
+
+SCOPE_KEY = PROXY_CLIENTSIDE_CREDENTIAL_SCOPE_METADATA_KEY
+ATTACKER_API_BASE = "http://127.0.0.1:9/attacker"
+
+DEPLOYMENT_WITHOUT_OPT_IN = {
+    "model_name": "model-b",
+    "litellm_params": {
+        "model": "openai/gpt-4o",
+        "api_key": "sk-backend-b",
+        "api_base": "http://legit-upstream",
+    },
+}
+DEPLOYMENT_WITH_OPT_IN = {
+    "model_name": "model-b",
+    "litellm_params": {
+        "model": "openai/gpt-4o",
+        "api_key": "sk-backend-b",
+        "api_base": "http://legit-upstream",
+        "configurable_clientside_auth_params": ["api_base"],
+    },
+}
+
+
+def _kwargs_with_caller_credential() -> dict:
+    return {"api_base": ATTACKER_API_BASE, "messages": [{"role": "user", "content": "hi"}]}
+
+
+def test_caller_metadata_does_not_shadow_per_model_stamp_in_litellm_metadata():
+    """The reported bypass shape: unstamped caller ``metadata`` object + per_model
+    stamp in ``litellm_metadata`` (LITELLM_METADATA_ROUTES / thread routes)."""
+    kwargs = _kwargs_with_caller_credential()
+    kwargs["metadata"] = {"user_id": "legit-anthropic-metadata"}
+    kwargs["litellm_metadata"] = {SCOPE_KEY: "per_model"}
+
+    strip_clientside_credentials_without_deployment_opt_in(deployment=DEPLOYMENT_WITHOUT_OPT_IN, kwargs=kwargs)
+
+    assert "api_base" not in kwargs
+
+
+def test_reverse_bucket_order_still_strips():
+    """Stamp in ``metadata`` (regular routes) must not be shadowed by an
+    unstamped ``litellm_metadata`` bucket either."""
+    kwargs = _kwargs_with_caller_credential()
+    kwargs["metadata"] = {SCOPE_KEY: "per_model"}
+    kwargs["litellm_metadata"] = {"user_id": "unrelated"}
+
+    strip_clientside_credentials_without_deployment_opt_in(deployment=DEPLOYMENT_WITHOUT_OPT_IN, kwargs=kwargs)
+
+    assert "api_base" not in kwargs
+
+
+def test_json_string_metadata_bucket_does_not_shadow_stamp():
+    """A JSON-encoded string bucket (multipart/extra_body shape) cannot carry
+    the stamp and must not shadow the dict bucket that does."""
+    import json
+
+    kwargs = _kwargs_with_caller_credential()
+    kwargs["metadata"] = json.dumps({"user_id": "legit"})
+    kwargs["litellm_metadata"] = {SCOPE_KEY: "per_model"}
+
+    strip_clientside_credentials_without_deployment_opt_in(deployment=DEPLOYMENT_WITHOUT_OPT_IN, kwargs=kwargs)
+
+    assert "api_base" not in kwargs
+
+
+def test_conflicting_scopes_fail_closed_to_per_model():
+    """If buckets ever disagree, the most restrictive scope wins. (Caller-forged
+    scope values are stripped upstream, so this is defense in depth.)"""
+    kwargs = _kwargs_with_caller_credential()
+    kwargs["metadata"] = {SCOPE_KEY: "proxy_wide"}
+    kwargs["litellm_metadata"] = {SCOPE_KEY: "per_model"}
+
+    strip_clientside_credentials_without_deployment_opt_in(deployment=DEPLOYMENT_WITHOUT_OPT_IN, kwargs=kwargs)
+
+    assert "api_base" not in kwargs
+
+
+def test_bracket_encoded_and_case_variant_keys_do_not_shadow_stamp():
+    """Form-style ``metadata[...]`` keys and case variants are distinct kwargs
+    keys and must never be mistaken for (or shadow) a stamped bucket."""
+    kwargs = _kwargs_with_caller_credential()
+    kwargs["metadata[foo]"] = "bar"
+    kwargs["Metadata"] = {"user_id": "case-variant"}
+    kwargs["litellm_metadata"] = {SCOPE_KEY: "per_model"}
+
+    strip_clientside_credentials_without_deployment_opt_in(deployment=DEPLOYMENT_WITHOUT_OPT_IN, kwargs=kwargs)
+
+    assert "api_base" not in kwargs
+
+
+def test_no_scope_stamp_leaves_sdk_router_behavior_unchanged():
+    """Plain SDK completion call (no proxy stamp anywhere): the per-call
+    clientside credential feature must keep working."""
+    kwargs = _kwargs_with_caller_credential()
+    kwargs["metadata"] = {"user_id": "sdk-user"}
+
+    strip_clientside_credentials_without_deployment_opt_in(deployment=DEPLOYMENT_WITHOUT_OPT_IN, kwargs=kwargs)
+
+    assert kwargs["api_base"] == ATTACKER_API_BASE
+
+
+def test_missing_metadata_entirely_leaves_behavior_unchanged():
+    kwargs = _kwargs_with_caller_credential()
+
+    strip_clientside_credentials_without_deployment_opt_in(deployment=DEPLOYMENT_WITHOUT_OPT_IN, kwargs=kwargs)
+
+    assert kwargs["api_base"] == ATTACKER_API_BASE
+
+
+def test_proxy_wide_scope_leaves_credential_untouched():
+    """Admin opted every deployment in via
+    general_settings.allow_client_side_credentials."""
+    kwargs = _kwargs_with_caller_credential()
+    kwargs["metadata"] = {SCOPE_KEY: "proxy_wide"}
+
+    strip_clientside_credentials_without_deployment_opt_in(deployment=DEPLOYMENT_WITHOUT_OPT_IN, kwargs=kwargs)
+
+    assert kwargs["api_base"] == ATTACKER_API_BASE
+
+
+def test_opted_in_deployment_keeps_caller_credential_under_per_model():
+    """The benign primary-path case: the dispatched deployment opted in to the
+    credential key, so the caller's value is honored even under per_model."""
+    kwargs = _kwargs_with_caller_credential()
+    kwargs["metadata"] = {"user_id": "legit"}
+    kwargs["litellm_metadata"] = {SCOPE_KEY: "per_model"}
+
+    strip_clientside_credentials_without_deployment_opt_in(deployment=DEPLOYMENT_WITH_OPT_IN, kwargs=kwargs)
+
+    assert kwargs["api_base"] == ATTACKER_API_BASE