fix(guardrails): drop safeguard verdicts and explanations for rewritten tool_use inputs

The non-streaming /v1/messages write-back replaced a tool_use input but left the
safeguard_results the classifier stamped on the original input, so an explanation
could still quote the secret the guardrail masked. The response now keeps verdicts
only for unchanged tool_use blocks and drops every explanation, the rule the
ended-stream path already follows
This commit is contained in:
Jesus Velarde 2026-11-13 16:19:08 -08:00
parent f394b968c5
commit 06a6195a81
2 changed files with 72 additions and 9 deletions

View file

@ -468,6 +468,18 @@ def _write_back_tool_use_block(block: object, shape: _ToolCallShape, rewritten_i
block.name = shape.name # rebind-ok: same in-place rewrite as the input above
def _redact_output_safeguard_results(response: object, unchanged_tool_use_ids: frozenset[str]) -> None:
"""The response's safeguard verdicts judged the tool inputs as the model wrote them, so only those
for unchanged ``tool_use`` blocks survive, and no explanation can quote an input the guardrail rewrote."""
fields: Final = as_str_mapping(response)
raw: Final = fields.get("safeguard_results") if fields is not None else None
if raw is None or not isinstance(response, dict):
return
response["safeguard_results"] = filtered_safeguard_results( # rebind-ok: client-visible response rewrite
raw, unchanged_tool_use_ids, keep_explanations=False
)
def _streamed_query_strings(
tool_call_blocks: Sequence[tuple[int, Mapping[str, object]]], tool_calls: Sequence[_ToolCallShape]
) -> tuple[_WebSearchString, ...]:
@ -1558,13 +1570,15 @@ class AnthropicMessagesHandler(BaseTranslation):
self._write_back_web_search_strings(
response_content, _web_search_rewrites_by_block(web_search_strings, web_search_rewrites)
)
self._write_back_output_tool_calls(
unchanged_tool_use_ids: Final = self._write_back_output_tool_calls(
response_content,
scanned_tool_calls=tool_calls_to_check,
pre_guardrail_tool_calls=pre_guardrail_tool_calls,
returned_tool_calls=guardrailed_inputs.get("tool_calls"),
guardrail_name=guardrail_to_apply.guardrail_name,
)
if unchanged_tool_use_ids is not None:
_redact_output_safeguard_results(response, unchanged_tool_use_ids)
verbose_proxy_logger.debug("Anthropic Messages: Processed output response: %s", response)
@ -1750,33 +1764,42 @@ class AnthropicMessagesHandler(BaseTranslation):
pre_guardrail_tool_calls: tuple[_ToolCallShape, ...],
returned_tool_calls: Sequence[object] | None,
guardrail_name: str | None,
) -> None:
) -> frozenset[str] | None:
"""The nth returned call belongs to the nth ``tool_use`` block, the order the calls were scanned
in. A returned list of another length cannot be placed, so the scanned calls stand, as on the
request side. Rewritten arguments that are not a JSON object cannot become a ``tool_use`` input,
so the response fails instead of carrying the arguments the guardrail meant to replace."""
so the response fails instead of carrying the arguments the guardrail meant to replace. Returns the
ids of the ``tool_use`` blocks left unchanged when any block was rewritten, else None."""
post_guardrail_tool_calls: Final = _tool_call_shapes(
returned_tool_calls
if returned_tool_calls is not None and len(returned_tool_calls) == len(pre_guardrail_tool_calls)
else scanned_tool_calls
)
tool_use_blocks: Final = tuple(
block
(block, fields[1].get("id"))
for block in response_content
if (fields := cls._output_block_fields(block)) is not None and fields[0] == "tool_use"
)
if len(tool_use_blocks) != len(pre_guardrail_tool_calls):
return
return None
rewrites: Final = tuple(
(block, after, _rewritten_tool_use_input(after.arguments))
for block, before, after in zip(tool_use_blocks, pre_guardrail_tool_calls, post_guardrail_tool_calls)
(position, after, _rewritten_tool_use_input(after.arguments))
for position, (before, after) in enumerate(zip(pre_guardrail_tool_calls, post_guardrail_tool_calls))
if before != after
)
if not rewrites:
return None
if any(rewritten_input is None for _, _, rewritten_input in rewrites):
raise unappliable_request_rewrite(guardrail_name)
for block, after, rewritten_input in rewrites:
for position, after, rewritten_input in rewrites:
if rewritten_input is not None:
_write_back_tool_use_block(block, after, rewritten_input)
_write_back_tool_use_block(tool_use_blocks[position][0], after, rewritten_input)
rewritten_positions: Final = frozenset(position for position, _, _ in rewrites)
return frozenset(
tool_use_id
for position, (_, tool_use_id) in enumerate(tool_use_blocks)
if position not in rewritten_positions and isinstance(tool_use_id, str)
)
@staticmethod
def _get_response_content(response: object) -> Sequence[object]:

View file

@ -3659,6 +3659,46 @@ class TestOutputToolCallRewrites:
"[SCANNED] and then",
]
@pytest.mark.asyncio
async def test_a_rewritten_call_drops_its_safeguard_verdict_and_every_explanation(self) -> None:
response: Final = _dict_tool_use_response()
content: Final = cast(list[object], response["content"])
content.append({"type": "tool_use", "id": "toolu_2", "name": "lookup", "input": {"query": "weather"}})
response["safeguard_results"] = [
{
"type": "dangerous_tool_use",
"status": {
"type": "available",
"explanation": "runs echo SECRET-123",
"tool_uses": {
"toolu_1": {"type": "evaluated", "outcome": "flagged", "explanation": "echoes SECRET-123"},
"toolu_2": {"type": "evaluated", "outcome": "not_flagged", "explanation": "looks up weather"},
},
},
}
]
def redact_first_only(tool_calls: Sequence[object]) -> list[object]:
return [*_redacting(tool_calls[:1]), tool_calls[1]]
await AnthropicMessagesHandler().process_output_response(
response=cast(Any, response),
guardrail_to_apply=_tool_call_returning_guardrail(redact_first_only),
litellm_logging_obj=MagicMock(),
)
assert _tool_use_block(response)["input"] == {"cmd": "echo [REDACTED]"}
assert response["safeguard_results"] == (
{
"type": "dangerous_tool_use",
"status": {
"type": "available",
"tool_uses": {"toolu_2": {"type": "evaluated", "outcome": "not_flagged"}},
},
},
)
assert "SECRET-123" not in json.dumps(response)
@pytest.mark.asyncio
async def test_rewritten_arguments_that_are_not_a_json_object_fail_the_response(self) -> None:
response: Final = _dict_tool_use_response()