From f06c6c94ae6dc77eb1e17a8929f9d2a531d274f7 Mon Sep 17 00:00:00 2001 From: Caduri Katzav Date: Mon, 28 Sep 2026 19:09:19 +0300 Subject: [PATCH] fix(guardrails): keep masked text when a guardrail echoes multipart rows The request model validates list message content lazily, so the dump that builds the POST body consumes it and a later read of the model rows sees empty content. apply_guardrail compared the guardrail's returned rows with those consumed rows, so no multipart row ever matched its echo. A guardrail that echoed every row unchanged and masked through texts had its rows taken as a rewrite, and the unmasked rows reached the LLM Compare against the JSON rows actually posted instead. An unchanged echo now falls back to texts again, and a partial echo restores the caller's row at each echoed index. structured_messages_from_response is renamed to structured_messages_from_json since it now reads request rows as well The dump also sends a row as content [] when one of its parts fails validation (guarded_text, a base64 document, flac audio), so the guardrail never saw that row's text. An echo of it restored the caller's unmasked row while another row's rewrite made the rows path win, which dropped the masked texts. Such a row now carries the caller's own content, so the guardrail sees it in full. Rows the model dumps in full are posted exactly as before, and the rest of the body is unchanged That content goes through a new as_json_value helper in _content_utils. It round-trips through the stdlib json codec because pydantic's serializer silently replaces anything nested past 254 levels with "..." GenericGuardrailAPI also takes an optional async_handler so tests can inject the HTTP client. The regression tests drive the real OpenAI chat and Anthropic Messages guardrail handlers through it --- litellm/proxy/guardrails/_content_utils.py | 13 + .../generic_guardrail_api.py | 42 ++- .../guardrail_hooks/generic_guardrail_api.py | 4 +- .../test_generic_guardrail_api.py | 290 ++++++++++++++++++ .../proxy/guardrails/test_content_utils.py | 12 + 5 files changed, 354 insertions(+), 7 deletions(-) diff --git a/litellm/proxy/guardrails/_content_utils.py b/litellm/proxy/guardrails/_content_utils.py index 7529fe99f52..44142245390 100644 --- a/litellm/proxy/guardrails/_content_utils.py +++ b/litellm/proxy/guardrails/_content_utils.py @@ -8,9 +8,13 @@ skip the other shapes — these helpers normalise that so every hook sees every text fragment. """ +import json from collections.abc import Callable, Iterator, Mapping, Sequence from typing import Any, Final +from pydantic import JsonValue +from pydantic_core import to_jsonable_python + # Call types whose body carries free-form chat / prompt text that # text-content guardrails (banned keywords, content moderation, secret # detection, …) should inspect. The proxy ingress passes ``route_type`` @@ -307,3 +311,12 @@ def build_inspection_messages(data: dict[str, Any]) -> list[dict[str, str]]: role = message.get("role", "user") or "user" flattened.append({"role": role, "content": text}) return flattened + + +def as_json_value(value: object) -> JsonValue: + """Round-trips through the stdlib codec because pydantic's serializer turns anything nested + past 254 levels into "...", while this keeps about the depth the proxy's request parser accepts""" + parsed: Final[JsonValue] = json.loads( # pyright: ignore[reportAny] # untyped stdlib parse of json.dumps output + json.dumps(value, default=to_jsonable_python) + ) + return parsed diff --git a/litellm/proxy/guardrails/guardrail_hooks/generic_guardrail_api/generic_guardrail_api.py b/litellm/proxy/guardrails/guardrail_hooks/generic_guardrail_api/generic_guardrail_api.py index 3d1a173635e..db4b05a158f 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/generic_guardrail_api/generic_guardrail_api.py +++ b/litellm/proxy/guardrails/guardrail_hooks/generic_guardrail_api/generic_guardrail_api.py @@ -11,6 +11,8 @@ from collections.abc import Mapping, Sequence from typing import TYPE_CHECKING, Any, Final, Literal, Optional import httpx +from pydantic import JsonValue +from typing_extensions import TypeIs from litellm._logging import verbose_proxy_logger from litellm._version import version as litellm_version @@ -20,9 +22,11 @@ from litellm.integrations.custom_guardrail import ( log_guardrail_information, ) from litellm.llms.custom_httpx.http_handler import ( + AsyncHTTPHandler, get_async_httpx_client, httpxSpecialProvider, ) +from litellm.proxy.guardrails._content_utils import as_json_value from litellm.types.guardrails import GuardrailEventHooks from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam from litellm.types.proxy.guardrails.guardrail_hooks.generic_guardrail_api import ( @@ -30,6 +34,7 @@ from litellm.types.proxy.guardrails.guardrail_hooks.generic_guardrail_api import GenericGuardrailAPIRequest, GenericGuardrailAPIResponse, GuardrailToolParam, + structured_messages_from_json, ) from litellm.types.utils import GenericGuardrailAPIInputs @@ -150,6 +155,28 @@ def _extract_inbound_headers( return None +def _is_part_list(value: object) -> TypeIs[list[object]]: # guard-ok: trivial isinstance narrowing + return isinstance(value, list) + + +def _row_as_sent(dumped: JsonValue, caller: Mapping[str, object]) -> JsonValue: + """The request model validates list content lazily and dumps a list holding + any part it rejects as [], so such a row is sent with the caller's content.""" + caller_content: Final = caller.get("content") + if not isinstance(dumped, dict) or not _is_part_list(caller_content): + return dumped + dumped_content: Final = dumped.get("content") + if isinstance(dumped_content, list) and len(dumped_content) == len(caller_content): + return dumped + return {**dumped, "content": as_json_value(caller_content)} + + +def _rows_as_sent(dumped_rows: JsonValue, caller_rows: Sequence[Mapping[str, object]] | None) -> JsonValue: + if caller_rows is None or not isinstance(dumped_rows, list): + return dumped_rows + return [_row_as_sent(dumped, caller) for dumped, caller in zip(dumped_rows, caller_rows, strict=True)] + + def _structured_rows_to_write_back( original_rows: Sequence[AllMessageValues] | None, shown_rows: Sequence[AllMessageValues] | None, @@ -204,9 +231,12 @@ class GenericGuardrailAPI(CustomGuardrail): streaming_end_of_stream_only: bool | None = None, streaming_sampling_rate: int | None = None, streaming_transform_mode: Literal["block_only", "incremental_diff"] | None = None, + async_handler: AsyncHTTPHandler | None = None, **kwargs, ): - self.async_handler = get_async_httpx_client(llm_provider=httpxSpecialProvider.GuardrailCallback) + self.async_handler = async_handler or get_async_httpx_client( + llm_provider=httpxSpecialProvider.GuardrailCallback + ) self.headers = headers or {} self.extra_headers = extra_headers or [] @@ -470,12 +500,14 @@ class GenericGuardrailAPI(CustomGuardrail): ) headers: Final = self._build_request_headers() + # The model's list content is a lazy iterator that this dump consumes, so it cannot be read again + dumped: Final[Mapping[str, JsonValue]] = guardrail_request.model_dump(mode="json") + sent_messages: Final = _rows_as_sent(dumped.get("structured_messages"), structured_messages) + request_json: Final = {**dumped, "structured_messages": sent_messages} # mutable-ok: JSON POST body - # Make the API request - # Use mode="json" to ensure all iterables are converted to lists response: Final = await self.async_handler.post( url=self.api_base, - json=guardrail_request.model_dump(mode="json"), + json=request_json, headers=headers, ) @@ -503,7 +535,7 @@ class GenericGuardrailAPI(CustomGuardrail): images=images, tools=tools, structured_messages=structured_messages, - shown_messages=guardrail_request.structured_messages, + shown_messages=structured_messages_from_json(sent_messages), guardrail_response=guardrail_response, ) diff --git a/litellm/types/proxy/guardrails/guardrail_hooks/generic_guardrail_api.py b/litellm/types/proxy/guardrails/guardrail_hooks/generic_guardrail_api.py index 44e2cc2404f..c199bfe91b0 100644 --- a/litellm/types/proxy/guardrails/guardrail_hooks/generic_guardrail_api.py +++ b/litellm/types/proxy/guardrails/guardrail_hooks/generic_guardrail_api.py @@ -159,7 +159,7 @@ def coerce_stream_holdback_value(value: Any) -> int: return 0 -def structured_messages_from_response(value: object) -> Sequence[AllMessageValues] | None: +def structured_messages_from_json(value: object) -> Sequence[AllMessageValues] | None: if not isinstance(value, list): return None if not all(isinstance(message, Mapping) and isinstance(message.get("role"), str) for message in value): @@ -212,5 +212,5 @@ class GenericGuardrailAPIResponse: images=data.get("images"), tools=data.get("tools"), stream_holdback_chars=stream_holdback_chars, - structured_messages=structured_messages_from_response(data.get("structured_messages")), + structured_messages=structured_messages_from_json(data.get("structured_messages")), ) diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_generic_guardrail_api.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_generic_guardrail_api.py index a5e79f84ef1..7cc958697cf 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_generic_guardrail_api.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_generic_guardrail_api.py @@ -5,16 +5,23 @@ This test file tests the Generic Guardrail API implementation, specifically focusing on metadata extraction and passing. """ +import json import os +from collections.abc import Callable, Mapping +from typing import Final from unittest.mock import AsyncMock, MagicMock, patch import httpx import pytest +from pydantic import JsonValue import litellm from litellm import ModelResponse from litellm._version import version as litellm_version from litellm.exceptions import GuardrailRaisedException, Timeout +from litellm.llms.anthropic.chat.guardrail_translation.handler import AnthropicMessagesHandler +from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler +from litellm.llms.openai.chat.guardrail_translation.handler import OpenAIChatCompletionsHandler from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.guardrails.guardrail_hooks.generic_guardrail_api import ( GenericGuardrailAPI, @@ -22,6 +29,8 @@ from litellm.proxy.guardrails.guardrail_hooks.generic_guardrail_api import ( from litellm.proxy.guardrails.guardrail_hooks.generic_guardrail_api.generic_guardrail_api import ( _HEADER_PRESENT_PLACEHOLDER, ) +from litellm.types.llms.anthropic import AllAnthropicMessageValues +from litellm.types.llms.openai import AllMessageValues, ChatCompletionImageObject from litellm.types.utils import Choices, Message @@ -721,6 +730,287 @@ class TestStructuredMessagesInResponse: assert guardrailed_inputs["texts"] == ["[REDACTED]"] +_SSN: Final = "123-45-6789" + + +def _image_part() -> ChatCompletionImageObject: + return {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo="}} + + +GuardrailAnswer = Callable[[Mapping[str, JsonValue]], Mapping[str, JsonValue]] + + +def _guardrail_answering(answer: GuardrailAnswer) -> GenericGuardrailAPI: + def serve(request: httpx.Request) -> httpx.Response: + return httpx.Response(200, json=answer(json.loads(request.content))) + + return GenericGuardrailAPI( + api_base="https://guardrail.test/beta/litellm_basic_guardrail_api", + guardrail_name="pii-masker", + event_hook="pre_call", + default_on=True, + async_handler=AsyncHTTPHandler(transport=httpx.MockTransport(serve)), + ) + + +def _masked(text: str) -> str: + return text.replace(_SSN, "[SSN]") + + +def _echo_every_row_and_mask_texts(request_json: Mapping[str, JsonValue]) -> Mapping[str, JsonValue]: + return { + "action": "GUARDRAIL_INTERVENED", + "structured_messages": request_json["structured_messages"], + "texts": [_masked(text) for text in request_json["texts"]], + } + + +def _echo_first_row_and_mask_the_rest(request_json: Mapping[str, JsonValue]) -> Mapping[str, JsonValue]: + first_row, *other_rows = request_json["structured_messages"] + return { + "action": "GUARDRAIL_INTERVENED", + "structured_messages": [first_row, *({**row, "content": _masked(row["content"])} for row in other_rows)], + } + + +def _masked_part(part: Mapping[str, JsonValue]) -> Mapping[str, JsonValue]: + return {**part, "text": _masked(part["text"])} if part["type"] == "text" else part + + +def _masked_row(row: Mapping[str, JsonValue]) -> Mapping[str, JsonValue]: + content: Final = row["content"] + if isinstance(content, str): + return {**row, "content": _masked(content)} + return {**row, "content": [_masked_part(part) for part in content]} + + +def _mask_every_row_and_text(request_json: Mapping[str, JsonValue]) -> Mapping[str, JsonValue]: + return { + "action": "GUARDRAIL_INTERVENED", + "texts": [_masked(text) for text in request_json["texts"]], + "structured_messages": [_masked_row(row) for row in request_json["structured_messages"]], + } + + +def _guarded_text_part() -> Mapping[str, JsonValue]: + return {"type": "guarded_text", "text": "keep this guarded"} + + +def _pdf_document_part() -> Mapping[str, JsonValue]: + return {"type": "document", "source": {"type": "base64", "media_type": "application/pdf", "data": "JVBERi0="}} + + +def _nested(depth: int) -> JsonValue: + return {"leaf": "x"} if depth == 0 else {"nested": _nested(depth - 1)} + + +async def _llm_bound_messages(guardrail: GenericGuardrailAPI, messages: list[AllMessageValues]) -> object: + data: Final = await OpenAIChatCompletionsHandler().process_input_messages( + data={"model": "gpt-5.6", "messages": messages}, guardrail_to_apply=guardrail + ) + return data["messages"] + + +async def _structured_messages_posted_for(rows: list[AllMessageValues]) -> list[JsonValue]: + posted: Final[list[JsonValue]] = [] # mutable-ok: records what the endpoint received + + def record(request_json: Mapping[str, JsonValue]) -> Mapping[str, JsonValue]: + posted.append(request_json["structured_messages"]) + return {"action": "NONE"} + + await _guardrail_answering(record).apply_guardrail( + inputs={"texts": ["hi"], "structured_messages": rows}, request_data={}, input_type="request" + ) + return posted + + +async def _llm_bound_anthropic_messages( + guardrail: GenericGuardrailAPI, messages: list[AllAnthropicMessageValues] +) -> object: + data: Final = await AnthropicMessagesHandler().process_input_messages( + data={"model": "claude-opus-5-5", "max_tokens": 64, "messages": messages}, guardrail_to_apply=guardrail + ) + return data["messages"] + + +class TestEchoedRowsReachingTheLLM: + @pytest.mark.asyncio + @pytest.mark.parametrize( + ("messages", "expected"), + [ + ( + [{"role": "user", "content": [{"type": "text", "text": f"my ssn is {_SSN}"}, _image_part()]}], + [{"role": "user", "content": [{"type": "text", "text": "my ssn is [SSN]"}, _image_part()]}], + ), + ( + [ + {"role": "system", "content": "You are helpful."}, + {"role": "user", "content": [{"type": "text", "text": f"ssn {_SSN}"}, _image_part()]}, + {"role": "user", "content": f"again {_SSN}"}, + ], + [ + {"role": "system", "content": "You are helpful."}, + {"role": "user", "content": [{"type": "text", "text": "ssn [SSN]"}, _image_part()]}, + {"role": "user", "content": "again [SSN]"}, + ], + ), + ( + [{"role": "user", "content": f"my ssn is {_SSN}"}], + [{"role": "user", "content": "my ssn is [SSN]"}], + ), + ], + ids=["multipart", "multipart_among_string_rows", "string_only"], + ) + async def test_every_row_echoed_applies_the_masked_texts( + self, messages: list[AllMessageValues], expected: list[AllMessageValues] + ) -> None: + guardrail: Final = _guardrail_answering(_echo_every_row_and_mask_texts) + + llm_bound: Final = await _llm_bound_messages(guardrail, messages) + + assert llm_bound == expected, "an unchanged echo of every row must leave the rewrite to texts" + + @pytest.mark.asyncio + async def test_every_anthropic_content_block_row_echoed_applies_the_masked_texts(self) -> None: + guardrail: Final = _guardrail_answering(_echo_every_row_and_mask_texts) + + llm_bound: Final = await _llm_bound_anthropic_messages( + guardrail, + [ + { + "role": "user", + "content": [{"type": "text", "text": f"my ssn is {_SSN}"}, {"type": "text", "text": "ok"}], + } + ], + ) + + assert llm_bound == [ + {"role": "user", "content": [{"type": "text", "text": "my ssn is [SSN]"}, {"type": "text", "text": "ok"}]} + ], "an unchanged echo of every content block row must leave the rewrite to texts" + + @pytest.mark.asyncio + async def test_an_echoed_multipart_row_is_restored_to_the_callers_row(self) -> None: + guardrail: Final = _guardrail_answering(_echo_first_row_and_mask_the_rest) + + llm_bound: Final = await _llm_bound_messages( + guardrail, + [ + {"role": "user", "name": "pat", "content": [{"type": "text", "text": "what is this?"}, _image_part()]}, + {"role": "user", "content": f"my ssn is {_SSN}"}, + ], + ) + + assert llm_bound == [ + {"role": "user", "name": "pat", "content": [{"type": "text", "text": "what is this?"}, _image_part()]}, + {"role": "user", "content": "my ssn is [SSN]"}, + ], "the echoed row must keep the caller's keys the request model drops" + + @pytest.mark.asyncio + @pytest.mark.parametrize( + "unvalidated_part", [_guarded_text_part, _pdf_document_part], ids=["guarded_text", "pdf_document"] + ) + async def test_a_row_holding_a_part_the_request_model_rejects_is_masked( + self, unvalidated_part: Callable[[], Mapping[str, JsonValue]] + ) -> None: + guardrail: Final = _guardrail_answering(_mask_every_row_and_text) + + llm_bound: Final = await _llm_bound_messages( + guardrail, + [ + {"role": "user", "content": [{"type": "text", "text": f"my ssn is {_SSN}"}, unvalidated_part()]}, + {"role": "user", "content": f"also {_SSN}"}, + ], + ) + + assert llm_bound == [ + {"role": "user", "content": [{"type": "text", "text": "my ssn is [SSN]"}, unvalidated_part()]}, + {"role": "user", "content": "also [SSN]"}, + ], "the guardrail must see the whole row so its masking of it reaches the LLM" + + @pytest.mark.asyncio + async def test_an_echoed_row_holding_a_part_the_request_model_rejects_is_restored_to_the_callers_row( + self, + ) -> None: + guardrail: Final = _guardrail_answering(_echo_first_row_and_mask_the_rest) + + llm_bound: Final = await _llm_bound_messages( + guardrail, + [ + {"role": "user", "name": "pat", "content": [{"type": "text", "text": "hi"}, _guarded_text_part()]}, + {"role": "user", "content": f"my ssn is {_SSN}"}, + ], + ) + + assert llm_bound == [ + {"role": "user", "name": "pat", "content": [{"type": "text", "text": "hi"}, _guarded_text_part()]}, + {"role": "user", "content": "my ssn is [SSN]"}, + ], "the echoed row must keep the caller's keys the request model drops" + + @pytest.mark.asyncio + async def test_rows_the_request_model_accepts_are_posted_as_it_dumps_them(self) -> None: + posted: Final = await _structured_messages_posted_for( + [ + { + "role": "user", + "name": "pat", + "content": [ + {"type": "text", "text": "hi"}, + {**_image_part(), "cache_control": {"type": "ephemeral"}}, + ], + }, + { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_1", + "type": "function", + "function": {"name": "lookup", "arguments": "{}"}, + "index": 0, + } + ], + }, + {"role": "tool", "tool_call_id": "call_1", "content": "done"}, + ] + ) + + assert posted == [ + [ + {"role": "user", "content": [{"type": "text", "text": "hi"}, _image_part()]}, + { + "role": "assistant", + "content": None, + "tool_calls": [ + {"id": "call_1", "type": "function", "function": {"name": "lookup", "arguments": "{}"}} + ], + }, + {"role": "tool", "tool_call_id": "call_1", "content": "done"}, + ] + ], "rows the request model dumps in full must reach the guardrail exactly as before" + + @pytest.mark.asyncio + async def test_a_row_holding_a_part_the_request_model_rejects_is_posted_with_the_callers_content(self) -> None: + posted: Final = await _structured_messages_posted_for( + [{"role": "user", "name": "pat", "content": [{"type": "text", "text": "hi"}, _pdf_document_part()]}] + ) + + assert posted == [[{"role": "user", "content": [{"type": "text", "text": "hi"}, _pdf_document_part()]}]], ( + "only the content the request model emptied is taken from the caller, the row keys stay as dumped" + ) + + @pytest.mark.asyncio + async def test_a_deeply_nested_part_the_proxy_accepts_is_posted_in_full(self) -> None: + deep_document: Final = {"type": "document", "source": _nested(300)} + + posted: Final = await _structured_messages_posted_for( + [{"role": "user", "content": [{"type": "text", "text": "hi"}, deep_document]}] + ) + + assert posted == [[{"role": "user", "content": [{"type": "text", "text": "hi"}, deep_document]}]], ( + "content nested as deep as the proxy's own JSON parser allows must still reach the guardrail" + ) + + class TestImageSupport: """Test image handling in guardrail requests""" diff --git a/tests/test_litellm/proxy/guardrails/test_content_utils.py b/tests/test_litellm/proxy/guardrails/test_content_utils.py index 920ffc77095..527e0f39c94 100644 --- a/tests/test_litellm/proxy/guardrails/test_content_utils.py +++ b/tests/test_litellm/proxy/guardrails/test_content_utils.py @@ -2,6 +2,7 @@ from litellm.proxy.guardrails._content_utils import ( apply_redacted_messages_back, + as_json_value, build_inspection_messages, has_non_string_content, is_non_conversational_call_type, @@ -741,3 +742,14 @@ def test_is_non_conversational_call_type_defaults_to_inspecting_unknown_call_typ """A call type this module has never heard of must still be inspected — failing closed is the point of the deny-list.""" assert is_non_conversational_call_type("some_future_call_type") is False + + +def _nested(depth: int) -> dict[str, object]: + return {"leaf": "x"} if depth == 0 else {"nested": _nested(depth - 1)} + + +def test_as_json_value_keeps_content_nested_past_the_pydantic_serializer_limit_and_decodes_bytes(): + assert as_json_value([{"type": "document", "source": _nested(600), "data": b"raw"}, ("a", 1)]) == [ + {"type": "document", "source": _nested(600), "data": "raw"}, + ["a", 1], + ], "nothing may be truncated to '...' and non-JSON types must become their JSON form"