adversaria 0.1.0

Adversarial Testing Harness for Large Language Models

Adversaria Python SDK

Python SDK for Adversaria - Adversarial Testing Harness for LLMs.

Installation

pip install adversaria

Note: This SDK requires the Adversaria CLI to be installed:

cargo install adversaria

Quick Start

from adversaria import Adversaria

# Create client
client = Adversaria()

# Run tests
result = client.run(
    provider="openai",
    model="gpt-4",
    api_key="sk-..."  # Optional, uses OPENAI_API_KEY env var
)

# Check results
print(f"Risk Score: {result.risk_score}/100")
print(f"Successful Attacks: {result.successful_attacks}/{result.total_attacks}")

# Save report
report_path = result.save_report("./reports")
print(f"Report saved: {report_path}")

API Reference

Adversaria

Main client class for running security tests.

client = Adversaria(config_path=None)

Methods:

run(provider, model, suites=None, api_key=None)

Run security tests against an LLM.

Arguments:

• provider (str): Provider name ('openai', 'anthropic', 'ollama')
• model (str): Model name
• suites (list, optional): List of suite IDs to run
• api_key (str, optional): API key (uses env var if not provided)

Returns: TestResult

list_suites()

List available attack suites.

Returns: List[Suite]

list_reports(directory="./reports")

List available reports.

Returns: List[str] - List of report file paths

load_report(filepath)

Load a report from file.

Returns: TestResult

TestResult

Test execution results.

Attributes:

• id (str): Unique test run ID
• model (str): Model tested
• provider (str): Provider used
• timestamp (str): Test timestamp
• risk_score (int): Overall risk score (0-100)
• total_attacks (int): Total attacks executed
• successful_attacks (int): Number of successful attacks
• failed_attacks (int): Number of failed attacks
• duration_ms (int): Execution time in milliseconds
• raw_data (dict): Full report data

Methods:

save_report(directory="./reports")

Save report to directory.

Returns: str - Path to saved report

Examples

Basic Testing

import os
from adversaria import Adversaria

# Set API key
os.environ["OPENAI_API_KEY"] = "sk-..."

# Run test
client = Adversaria()
result = client.run(provider="openai", model="gpt-4")

if result.risk_score > 50:
    print("⚠️ High risk detected!")

Specific Suites

result = client.run(
    provider="openai",
    model="gpt-4",
    suites=["prompt_injection", "jailbreak"]
)

Multiple Providers

providers = [
    ("openai", "gpt-4"),
    ("anthropic", "claude-3-opus-20240229"),
]

for provider, model in providers:
    result = client.run(provider=provider, model=model)
    print(f"{provider}/{model}: {result.risk_score}/100")

Load Previous Report

client = Adversaria()

# List all reports
reports = client.list_reports()

# Load latest report
if reports:
    result = client.load_report(reports[0])
    print(f"Risk Score: {result.risk_score}/100")

Requirements

• Python 3.8+
• Adversaria CLI (cargo install adversaria)
• PyYAML

License

MIT