Skip to main content
This page provides detailed information about all configuration options for the EnkryptAI Red Teaming API V3 payload structure.

​
Payload Structure Overview

Every payload consists of three main sections: 
{
    "dataset_configuration": {
        // Dataset settings and parameters
    },
    "redteam_test_configurations": {
        // Test settings and parameters
    },
    "endpoint_configuration": {
        // Details of the target model
    }
}

​
Dataset Configuration

The dataset configuration specifies settings for generating test prompts when using generated datasets.

​
Complete Example

{
    "dataset_configuration": {
        "system_description": "System description for your AI application",
        "policy_description": "Do not generate any content that is illegal, harmful, or violates the rights of others.",
        "max_prompts": 100,
        "scenarios": 2,
        "categories": 2,
        "depth": 2
    }
}

​
Field Reference

​
system_description
string
required
Description of the AI system being tested. This helps the attack generator create contextually relevant prompts.Example: "This is a customer support chatbot for a banking application"
​
policy_description
string
required
The code of conduct or policy that the model should follow. Defines what the model should NOT do.Example: "Do not provide harmful or illegal information. Do not share customer data."
​
max_prompts
integer
default:"100"
Maximum number of prompts to generate for the dataset.Range: 1-1000Guidelines:
  • Development: 10-50
  • Standard testing: 100-200
  • Comprehensive: 500-1000
​
scenarios
integer
default:"2"
Number of different scenarios to test across.Range: 1-10Description: Creates variety in attack contexts and use cases.
​
categories
integer
default:"2"
Number of risk categories to include in the generated dataset.Range: 1-15Description: Spreads testing across different types of risks (toxicity, bias, PII, etc.).
​
depth
integer
default:"2"
Depth of the attack tree for generating adversarial prompts.Range: 1-5Description: Higher depth creates more sophisticated, multi-step attacks.Guidelines:
  • Depth 1-2: Simple attacks
  • Depth 3: Moderate complexity
  • Depth 4-5: Complex, multi-layered attacks

​
Usage Notes

The dataset configuration is only used when generating custom datasets. If you’re testing with pre-existing prompts, you can omit this section.

​
Red Team Test Configurations

This section specifies which tests to run and how to attack the model for each test type.

​
Structure

{
  "redteam_test_configurations": {
    "test_name": {
      "sample_percentage": 50,
      "attack_methods": {
        // Attack method configuration
      }
    }
  }
}

​
Common Fields

​
sample_percentage
integer
required
Percentage of the dataset to use for this specific test.Range: 1-100Guidelines:
  • Development/Quick Testing: 1-5%
  • Standard Testing: 10-20%
  • Comprehensive Testing: 50-100%
Example: "sample_percentage": 50 means use 50% of available prompts
​
attack_methods
object
required
Configuration of attack methods to use for this test. See Attack Methods Reference for detailed information.Structure:
{
  "basic": {"basic": {"params": {}}},
  "static": {
    "encoding_method": {"params": {...}}
  },
  "dynamic": {
    "iterative_method": {"params": {...}}
  }
}

​
Test Categories

​
custom_test

Custom dataset test for your specific use case.Example:
{
  "custom_test": {
    "sample_percentage": 50,
    "attack_methods": {
      "basic": {"basic": {"params": {}}}
    }
  }
}

​
bias_test

Identifies and exposes biased outputs based on protected attributes (race, gender, age, etc.).

​
cbrn_test

Addresses vulnerabilities related to chemical, biological, radiological, and nuclear domains.

​
csem_test

Addresses vulnerabilities related to child abuse exploitation.

​
harmful_test

Elicits responses that promote harm or danger to individuals or groups.

​
insecure_code_test

Produces insecure or harmful code snippets with vulnerabilities.

​
toxicity_test

Generates harmful or offensive content including hate speech and profanity.

​
pii_test

Exposes personally identifiable information that should remain private.Exposes copyrighted material that the model should not reproduce.

​
misinformation_test

Tests the model’s tendency to generate false or misleading information.

​
system_prompt_extractions_test

Attempts to extract the system prompt or instructions.

​
sponge_test

Tests for infinite loops or resource exhaustion attacks.

​
competitor_test

Attempts to extract information about competitors.Example:
{
  "bias_test": {
    "sample_percentage": 30,
    "attack_methods": {
      "basic": {"basic": {"params": {}}},
      "static": {"obfuscation": {"params": {}}}
    }
  }
}
Tests specifically designed for AI agents with tool use and autonomous capabilities:

​
governance_test

Testing the alignment and governance of the model including goal misalignment and policy drift.

​
agent_output_quality_test

Testing the quality of the agent’s output including hallucinations, bias, and toxicity in multi-turn workflows.

​
tool_misuse_test

Testing the misuse of the model’s tools including API integration issues, supply chain vulnerabilities, and resource consumption.

​
privacy_test

Testing privacy protections and data handling including sensitive data exposure and exfiltration channels.

​
reliability_and_observability_test

Testing the reliability and observability of agent operations including data poisoning, concept drift, and opaque reasoning.

​
agent_behaviour_test

Testing the behaviour patterns and decision-making including human manipulation and unsafe actuation.

​
access_control_and_permissions_test

Testing access control and permissions enforcement including credential theft, privilege escalation, and confused deputy attacks.

​
tool_extraction_test

Test if the agent tool information can be extracted in outputs.Example:
{
  "tool_misuse_test": {
    "sample_percentage": 20,
    "attack_methods": {
      "basic": {"basic": {"params": {}}},
      "dynamic": {"multi_turn": {"params": {}}}
    }
  },
  "access_control_and_permissions_test": {
    "sample_percentage": 25,
    "attack_methods": {
      "basic": {"basic": {"params": {}}},
      "static": {"obfuscation": {"params": {}}}
    }
  }
}
Advanced tests that require generated adversarial datasets:

​
adv_bias_test

Uncover biased outputs through adversarial methods.

​
adv_info_test

Extract sensitive or unintended information from a generated dataset.

​
adv_persona_test

Adversarial persona manipulation and identity spoofing attacks.

​
adv_command_test

Manipulate the model to execute unintended commands.

​
adv_pii_test

Expose personally identifiable information through adversarial techniques.

​
adv_competitor_test

Exploit vulnerabilities to gain an advantage over competitors.

​
adv_sponge_test

Advanced resource exhaustion and denial of service attacks.

​
adv_persona_test

Adversarial persona manipulation and identity spoofing attacks.Example:
{
  "adv_bias_test": {
    "sample_percentage": 50,
    "attack_methods": {
      "basic": {"basic": {"params": {}}},
      "static": {
        "base64_encoding": {
          "params": {"encoding_type": "base64", "iterations": 2}
        }
      }
    }
  }
}

​
Multiple Tests Example

You can run multiple tests in a single payload: 
{
  "redteam_test_configurations": {
    "bias_test": {
      "sample_percentage": 30,
      "attack_methods": {
        "basic": {"basic": {"params": {}}}
      }
    },
    "harmful_test": {
      "sample_percentage": 50,
      "attack_methods": {
        "basic": {"basic": {"params": {}}},
        "static": {"obfuscation": {"params": {}}}
      }
    },
    "pii_test": {
      "sample_percentage": 20,
      "attack_methods": {
        "basic": {"basic": {"params": {}}}
      }
    }
  }
}

​
Endpoint Configuration

This section provides specifics about the model to be tested.

​
Complete Example

{
    "endpoint_configuration": {
        "testing_for": "foundationModels",
        "model_name": "gpt-4o",
        "model_config": {
            "model_provider": "openai",
            "endpoint": {
                "scheme": "https",
                "host": "api.openai.com",
                "port": 443,
                "base_path": "/v1/chat/completions"
            },
            "auth_data": {
                "header_name": "Authorization",
                "header_prefix": "Bearer",
                "space_after_prefix": true
            },
            "apikeys": ["OPENAI_API_KEY"],
            "input_modalities": ["text"],
            "output_modalities": ["text"]
        }
    }
}

​
Top-Level Fields

​
testing_for
string
required
Identifies the AI system type being tested.Options:
  • "foundationModels" - Base language models (GPT, Claude, Gemini, etc.)
  • "chatbotsAndCopilots" - Chat interfaces and coding assistants
  • "agents" - Autonomous AI agents with tool use
  • "URL" - Custom API endpoints
​
model_name
string
required
The name or identifier for the model.Examples:
  • "gpt-4o"
  • "gpt-4o-mini"
  • "claude-3-5-sonnet-20241022"
  • "google/gemma-7b-it"
​
model_config
object
required
Configuration object for the model endpoint. See detailed fields below.

​
Model Config Fields

​
model_provider
string
required
The provider of the model.Supported Providers:
  • "openai" - OpenAI models
  • "anthropic" - Anthropic models
  • "together" - Together AI
  • "groq" - Groq
  • "google" - Google (Gemini)
  • "aws_bedrock" - AWS Bedrock
  • "azure" - Azure OpenAI
  • "huggingface" - HuggingFace models
For complete list, see GET Defaults API.
​
endpoint
object
required
Object containing endpoint details.Subfields:
  • scheme (string): Protocol scheme (e.g., "https")
  • host (string): API host (e.g., "api.openai.com")
  • port (integer): Port number (e.g., 443)
  • base_path (string): Base path for API endpoint (e.g., "/v1/chat/completions")
​
auth_data
object
required
Authentication configuration. Format varies by provider.Standard Header Auth (most providers):
{
  "header_name": "Authorization",
  "header_prefix": "Bearer",
  "space_after_prefix": true
}
AWS Bedrock:
{
  "aws_access_key_id": "YOUR_ACCESS_KEY",
  "aws_secret_access_key": "YOUR_SECRET_KEY"
}
Google Gemini (Query Param):
{
  "param_name": "key"
}
​
apikeys
array
required
Array of API keys for authentication.Example: ["OPENAI_API_KEY"]Note: Can provide multiple keys for load balancing.
​
input_modalities
array
required
Input types the model accepts.Options:
  • ["text"] - Text-only (LLMs)
  • ["text", "image"] - Text and images (VLMs)
  • ["text", "audio"] - Text and audio (ALMs)
​
output_modalities
array
required
Output types the model produces.Typical: ["text"]
​
system_prompt
string
Optional. Sets initial model behavior or instructions.Example: "You are a helpful customer support assistant."
​
rate_per_min
integer
Optional. Rate limit for API calls per minute.Example: 60
​
paths
object
Optional. Object defining API path mappings.Example:
{
  "completions": "/completions",
  "chat": "/chat/completions"
}
​
metadata
object
Optional. Provider-specific configuration options.Azure OpenAI:
{
  "azure_instance": "your-instance",
  "azure_api_version": "2023-05-15",
  "azure_deployment_id": "your-deployment"
}
AWS Bedrock:
{
  "bedrock.aws_region": "us-east-1"
}
Anthropic:
{
  "anthropic_version": "2023-06-01"
}
HuggingFace:
{
  "huggingface.use_cache": true,
  "huggingface.wait_for_model": true
}

​
Model-Type Specific Examples

{
  "endpoint_configuration": {
    "testing_for": "foundationModels",
    "model_name": "gpt-4o-mini",
    "model_config": {
      "model_provider": "openai",
      "endpoint": {
        "scheme": "https",
        "host": "api.openai.com",
        "port": 443,
        "base_path": "/v1/chat/completions"
      },
      "auth_data": {
        "header_name": "Authorization",
        "header_prefix": "Bearer",
        "space_after_prefix": true
      },
      "apikeys": ["OPENAI_API_KEY"],
      "input_modalities": ["text"],
      "output_modalities": ["text"]
    }
  }
}

​
Provider-Specific Notes

OpenAI

Standard bearer authentication with Authorization: Bearer YOUR_KEY

Anthropic

Requires anthropic_version in metadata and x-api-key header

AWS Bedrock

Uses AWS credentials instead of API keys. Requires region in metadata.

Google Gemini

Uses query parameter authentication with ?key=YOUR_KEY

Azure OpenAI

Requires Azure-specific metadata including instance and deployment ID

HuggingFace

Supports caching and wait-for-model options in metadata

​
Best Practices

Security Best Practices:
  • Never commit API keys to version control
  • Use environment variables for sensitive credentials
  • Rotate API keys regularly
  • Use separate keys for testing vs production
Configuration Tips:
  • Start with sample_percentage of 2-5% for initial testing
  • Gradually increase to 50-100% for comprehensive assessments
  • Use multiple test types to cover different risk categories
  • Match attack methods to your specific security concerns
⌘I