Implementation:BerriAI Litellm Secret Detection

Attribute	Value
Sources	enterprise/litellm_enterprise/enterprise_callbacks/secret_detection.py
Domains	Security, Guardrails, Enterprise Callbacks
Last Updated	2026-02-15 16:00 GMT

Overview

_ENTERPRISE_SecretDetection is a guardrail that scans LLM request messages, prompts, and inputs for secrets and API keys, automatically redacting detected secrets with [REDACTED] before the request reaches the LLM provider.

Description

This class extends CustomGuardrail and leverages the detect-secrets library to scan user content for over 90 types of secrets and API keys. It operates as a pre-call hook in the LiteLLM proxy pipeline, intercepting requests before they are forwarded to LLM providers.

The detection engine uses a comprehensive configuration of built-in and custom secret plugins covering providers such as AWS, Azure, GitHub, GitLab, Slack, Stripe, Twilio, OpenAI, Datadog, and many more. Custom plugins are loaded from a secrets_plugins directory adjacent to the module.

The guardrail processes three types of request data:

• messages -- Chat completion message content strings
• prompt -- Direct prompt strings or lists of prompt strings
• input -- Embedding/moderation input strings or lists

The guardrail name is hide_secrets and can be disabled per-key via the permissions field on the API key.

Usage

Import and register _ENTERPRISE_SecretDetection as a callback/guardrail when you need to prevent accidental leakage of secrets in LLM request payloads.

Code Reference

Source Location

enterprise/litellm_enterprise/enterprise_callbacks/secret_detection.py

Signature

class _ENTERPRISE_SecretDetection(CustomGuardrail):
    def __init__(self, detect_secrets_config: Optional[dict] = None, **kwargs): ...
    def scan_message_for_secrets(self, message_content: str) -> list: ...
    async def should_run_check(self, user_api_key_dict: UserAPIKeyAuth) -> bool: ...
    async def async_pre_call_hook(
        self,
        user_api_key_dict: UserAPIKeyAuth,
        cache: DualCache,
        data: dict,
        call_type: str,
    ): ...

Import

from litellm_enterprise.enterprise_callbacks.secret_detection import _ENTERPRISE_SecretDetection

I/O Contract

Inputs

Parameter	Type	Description
detect_secrets_config	Optional[dict]	Custom detect-secrets configuration. Falls back to built-in config with 90+ detectors.
data	dict	Request data containing messages, prompt, or input fields.
user_api_key_dict	UserAPIKeyAuth	API key metadata; used to check per-key guardrail permissions.
call_type	str	One of "completion", "embeddings", "image_generation", "moderation".

Outputs

Output	Type	Description
Modified data	None (in-place mutation)	Secret values in messages, prompt, and input fields are replaced with [REDACTED].
Detected secrets	list[dict]	Each secret has type (detector name) and value (the secret string).

Usage Examples

# Register as a callback in LiteLLM proxy config
litellm_settings:
  callbacks:
    - hide_secrets

# Programmatic usage
from litellm_enterprise.enterprise_callbacks.secret_detection import _ENTERPRISE_SecretDetection

detector = _ENTERPRISE_SecretDetection()
secrets = detector.scan_message_for_secrets("My API key is sk-abc123xyz")
# Returns: [{"type": "OpenAIApiKeyDetector", "value": "sk-abc123xyz"}]

Related Pages

• BerriAI_Litellm_LLM_Guard -- LLM Guard content moderation integration
• BerriAI_Litellm_Aporia_AI_Guardrail -- Aporia AI guardrail integration
• BerriAI_Litellm_Enterprise_Custom_Guardrail -- Tag-based guardrail mode control