fix(spend-logs): mask token-shaped secrets in stored request and response content

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
This commit is contained in:
Krrish Dholakia 2026-10-25 00:35:49 +00:00
parent 5171b3fbd8
commit 8df62ad5c7
3 changed files with 123 additions and 5 deletions

View file

@ -14,11 +14,13 @@ from litellm.rust_bridge import diagnostics
REDACTED: Final = "REDACTED"
_PEM_PRIVATE_KEY_BLOCK: Final = r"-----BEGIN[A-Z \-]*PRIVATE KEY-----[\s\S]*?-----END[A-Z \-]*PRIVATE KEY-----"
def _build_secret_patterns() -> "re.Pattern[str]":
patterns: Final[list[str]] = [
# PEM private key / certificate blocks
r"-----BEGIN[A-Z \-]*PRIVATE KEY-----[\s\S]*?-----END[A-Z \-]*PRIVATE KEY-----",
_PEM_PRIVATE_KEY_BLOCK,
# GCP OAuth2 access tokens (ya29.*)
r"\bya29\.[A-Za-z0-9_.~+/-]+",
# Credential %s formatting (space separator, no key= prefix)
@ -100,6 +102,37 @@ def redact_string(value: str) -> str:
return diagnostics.run(lambda native: native.redact_text(value), lambda: python_redact_string(value))
_TOKEN_SHAPES: Final = (
r"ya29\.[A-Za-z0-9_.~+/-]{20,}",
r"(?:AKIA|ASIA)[0-9A-Z]{16}\b",
rf"sk-[A-Za-z0-9\-_]{{{MINIMUM_CUSTOM_KEY_LENGTH - len('sk-')},}}",
r"(?:sk|rk)_(?:live|test)_[A-Za-z0-9]{24,}",
r"AIza[0-9A-Za-z\-_]{35}",
r"dapi[0-9a-f]{32}\b",
r"(?:ghp|gho|ghu|ghs|ghr)_[A-Za-z0-9]{36,}",
r"github_pat_[A-Za-z0-9_]{82}",
r"glpat-[A-Za-z0-9\-_]{20,}",
r"xox[abposr]-[A-Za-z0-9-]{10,}",
r"hf_[A-Za-z0-9]{34,}",
r"npm_[A-Za-z0-9]{36}\b",
r"pypi-AgEIcHlwaS5vcmc[A-Za-z0-9\-_]{50,}",
r"SG\.[A-Za-z0-9_-]{22}\.[A-Za-z0-9_-]{43}\b",
r"eyJ[A-Za-z0-9_-]{10,}\.eyJ[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]+",
)
# Tokens start at a word boundary or right after a JSON escape like "\n", since tool output is often serialized JSON
_TOKEN_START: Final = r"(?:(?<!\w)|(?<=\\[nrt]))"
_TOKEN_RE: Final = re.compile(f"{_PEM_PRIVATE_KEY_BLOCK}|{_TOKEN_START}(?:{'|'.join(_TOKEN_SHAPES)})")
def redact_tokens(value: str, replacement: str = REDACTED) -> str:
"""Replace substrings shaped like a credential (provider key prefixes, JWTs, PEM private keys).
Unlike redact_string() this never matches on surrounding words such as ``password:`` or ``Basic``,
so it is safe to run over free-form prompt and tool content.
"""
return _TOKEN_RE.sub(replacement, value)
_UNIX_SYSTEM_PATH: Final = r"/(?:etc|var|opt|usr|home|root|private|Users|tmp|mnt|srv)/[^\s'\"\)\]}>,]+"
_WINDOWS_DRIVE_PATH: Final = r"[A-Za-z]:\\[^\s'\"\)\]}>,]+"
_PRIVATE_OR_LOOPBACK_IPV4: Final = (

View file

@ -60,6 +60,7 @@ from litellm.litellm_core_utils.litellm_logging import (
)
from litellm.litellm_core_utils.ptu_pricing import azure_spillover
from litellm.litellm_core_utils.safe_json_dumps import safe_dumps, strip_null_bytes
from litellm.litellm_core_utils.secret_redaction import redact_tokens
from litellm.litellm_core_utils.sensitive_data_masker import SensitiveDataMasker
from litellm.llms.anthropic.common_utils import resolve_used_client_oauth_token
from litellm.proxy._types import (
@ -1204,7 +1205,7 @@ def _get_messages_for_spend_logs_payload(
messages: Final = standard_logging_payload.get("messages")
if messages is not None:
try:
return safe_dumps(messages)
return _without_tokens(safe_dumps(messages))
except Exception:
return "{}"
return "{}"
@ -1233,6 +1234,10 @@ def _is_spend_log_content(parent: Mapping[str, object], key: str) -> bool:
)
def _without_tokens(value: str) -> str:
return redact_tokens(value, REDACTED_BY_LITELM_STRING)
def _truncated_for_db(value: str, max_length: int) -> str:
if len(value) <= max_length:
return value
@ -1248,7 +1253,7 @@ def _truncated_for_db(value: str, max_length: int) -> str:
def _compaction_summary_for_db(summary: str) -> str:
return _truncated_for_db(
summary, max(_get_max_string_length_prompt_in_db(), MAX_STRING_LENGTH_COMPACTION_SUMMARY_IN_DB)
_without_tokens(summary), max(_get_max_string_length_prompt_in_db(), MAX_STRING_LENGTH_COMPACTION_SUMMARY_IN_DB)
)
@ -1277,7 +1282,7 @@ def _sanitize_request_body_for_spend_logs_payload(
At every nesting level, also strips keys listed in _SENSITIVE_REQUEST_BODY_KEYS (e.g. secret_fields,
which holds raw HTTP headers including Authorization tokens), and replaces string values under keys
SensitiveDataMasker classifies as credentials with REDACTED_BY_LITELM_STRING, except inside tool payloads
and logprobs.
and logprobs. Token-shaped substrings (e.g. a GitHub token in a tool result) are masked everywhere.
"""
if visited is None:
visited = set()
@ -1298,7 +1303,7 @@ def _sanitize_request_body_for_spend_logs_payload(
elif isinstance(value, list):
return [_sanitize_value(item, mask_credentials) for item in value]
elif isinstance(value, str):
return _truncated_for_db(value, max_string_length_prompt_in_db)
return _truncated_for_db(_without_tokens(value), max_string_length_prompt_in_db)
return value
def _stored_value(key: str, value: object) -> object:

View file

@ -1374,6 +1374,86 @@ def test_get_response_for_spend_logs_payload_strips_null_bytes(mock_should_store
assert json.loads(response_json)["content"] == "answerhere"
_GITHUB_TOKEN: Final = "ghp_" + "a1B2c3D4e5" * 3 + "F6g7H8"
_TOKEN_SHAPED_SECRETS: Final = (
_GITHUB_TOKEN,
"github_pat_" + "A1b2C3d4E5" * 8 + "Z9",
"sk-ant-api03-" + "Qw3rTy" * 7,
"AKIAIOSFODNN7EXAMPLE",
"xoxb-1234567890-0987654321-AbCdEfGhIjKl",
"glpat-AbCdEfGhIj1234567890",
"hf_" + "AbCdEfGhIj" * 3 + "KlMn",
"eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIxMjM0NTY3ODkwIn0.dozjgNryP4J3jVmNHl0w5N_XgL0n3I9PlFUP0THsR8U",
"-----BEGIN RSA PRIVATE KEY-----\nMIIEpAIBAAKCAQEA\n-----END RSA PRIVATE KEY-----",
)
@pytest.mark.parametrize("secret", _TOKEN_SHAPED_SECRETS)
@patch("litellm.proxy.spend_tracking.spend_tracking_utils.should_store_prompts_and_responses_in_spend_logs")
def test_stored_request_masks_a_token_returned_in_an_anthropic_tool_result(mock_should_store, secret: str):
mock_should_store.return_value = True
tool_output: Final = json.dumps({"stdout": f"export TOKEN={secret}\ndone"})
request_body: Final = {
"model": "claude-sonnet-5-5",
"messages": [
{"role": "user", "content": [{"type": "tool_result", "tool_use_id": "toolu_1", "content": tool_output}]}
],
}
stored: Final = json.dumps(_sanitize_request_body_for_spend_logs_payload(request_body))
assert secret not in stored
assert secret[:12] not in stored
assert "export TOKEN=REDACTED_BY_LITELM" in json.loads(json.loads(stored)["messages"][0]["content"][0]["content"])[
"stdout"
]
@patch("litellm.proxy.spend_tracking.spend_tracking_utils.should_store_prompts_and_responses_in_spend_logs")
def test_stored_response_masks_a_token_in_a_tool_call_argument(mock_should_store):
mock_should_store.return_value = True
payload: Final = cast(
StandardLoggingPayload,
{
"response": {
"content": [
{"type": "text", "text": "Cloning now"},
{"type": "tool_use", "id": "toolu_1", "name": "bash", "input": {"cmd": f"git clone https://{_GITHUB_TOKEN}@github.com/o/r"}},
]
}
},
)
stored: Final = json.loads(_get_response_for_spend_logs_payload(payload))
assert stored["content"][0]["text"] == "Cloning now"
assert stored["content"][1]["input"]["cmd"] == "git clone https://REDACTED_BY_LITELM@github.com/o/r"
def test_stored_request_masks_a_token_kept_in_the_head_of_a_truncated_string():
long_text: Final = f"{_GITHUB_TOKEN} " + "x" * 10_000
stored: Final = _sanitize_request_body_for_spend_logs_payload({"text": long_text}, max_string_length_prompt_in_db=200)
assert stored["text"].startswith("REDACTED_BY_LITELM x")
assert LITELLM_TRUNCATED_PAYLOAD_FIELD in stored["text"]
@pytest.mark.parametrize(
"text",
(
"Use basic auth with password: hunter2 and a Bearer header",
"The task-management-system and risk-assessment-framework docs",
"See https://github.com/BerriAI/litellm and ghp_short",
"A" * 5000,
),
)
def test_stored_request_keeps_text_that_only_resembles_credentials(text: str):
stored: Final = _sanitize_request_body_for_spend_logs_payload({"text": text}, max_string_length_prompt_in_db=10_000)
assert stored["text"] == text
@patch("litellm.proxy.spend_tracking.spend_tracking_utils.should_store_prompts_and_responses_in_spend_logs")
def test_get_response_for_spend_logs_payload_truncates_large_embedding(
mock_should_store,