DeepTeam is a framework to red team LLMs and AI agents.
from deepteam import red_team
from deepteam.vulnerabilities import *
from deepteam.attacks.single_turn import Base64
async def model_callback(input: str) -> str:
return input
def test_red_teaming():
custom_vulnerability = CustomVulnerability(
criteria="Criteria",
name="Name",
# custom_prompt="Custom prompt",
types=["type1", "type2"],
)
vulnerabilities = [
Bias(),
ChildProtection(),
Ethics(),
Fairness(),
Toxicity(),
Misinformation(),
IllegalActivity(),
PromptLeakage(),
PIILeakage(),
BFLA(),
BOLA(),
RBAC(),
DebugAccess(),
ShellInjection(),
SQLInjection(),
SSRF(),
IntellectualProperty(),
Competition(),
GraphicContent(),
PersonalSafety(),
CrossContextRetrieval(),
SystemReconnaissance(),
# Agentic vulnerabilities
GoalTheft(),
RecursiveHijacking(),
Robustness(),
ExcessiveAgency(),
IndirectInstruction(),
ToolOrchestrationAbuse(),
AgentIdentityAbuse(),
ToolMetadataPoisoning(),
UnexpectedCodeExecution(),
InsecureInterAgentCommunication(),
AutonomousAgentDrift(),
ExploitToolAgent(),
ExternalSystemAbuse(),
custom_vulnerability,
]
attacks = [Base64()]
risk_assessment = red_team(
model_callback=model_callback,
vulnerabilities=vulnerabilities,
attacks=attacks,
attacks_per_vulnerability_type=1,
max_concurrent=20,
ignore_errors=False,
)
assert risk_assessment is not None
assert len(risk_assessment.test_cases) > len(vulnerabilities) * len(attacks)