DeepTeam is a framework to red team LLMs and AI agents.
import pytest
from deepteam.frameworks import EUAIAct
from deepteam.red_teamer.risk_assessment import RiskAssessment
from deepteam.vulnerabilities import BaseVulnerability
from deepteam.attacks import BaseAttack
from deepteam.frameworks.risk_category import RiskCategory
from deepteam.frameworks.eu_ai_act.risk_categories import (
EU_AI_ACT_CATEGORIES,
)
from deepteam import red_team
EU_AI_ACT_DEFAULT_CATEGORIES = {
"subliminal_manipulation",
"exploitation_of_vulnerabilities",
"social_scoring",
"biometric_categorisation",
"remote_biometric_id_live",
"remote_biometric_id_post",
"biometric_id",
"critical_infrastructure",
"education",
"employment",
"essential_services",
"law_enforcement",
"migration_border",
"justice_democracy",
}
class TestEUAIAct:
def test_eu_ai_act_init(self):
"""Test that EU AI Act framework can be instantiated."""
framework = EUAIAct()
assert framework is not None
def test_eu_ai_act_name(self):
"""Test that EU AI Act framework has correct name."""
framework = EUAIAct()
assert (
framework.name
== framework.get_name()
== "EU Artificial Intelligence Act"
)
def test_eu_ai_act_description(self):
"""Test that EU AI Act framework has a non-empty description."""
framework = EUAIAct()
assert isinstance(framework.description, str)
assert len(framework.description) > 0
def test_eu_ai_act_default_categories(self):
"""Test that all EU AI Act categories are included by default."""
framework = EUAIAct()
assert set(framework.categories) == EU_AI_ACT_DEFAULT_CATEGORIES
def test_eu_ai_act_partial_categories(self):
"""Test that EU AI Act can be initialized with limited categories."""
framework = EUAIAct(categories=["social_scoring", "employment"])
assert set(framework.categories) == {
"social_scoring",
"employment",
}
assert len(framework.risk_categories) == 2
def test_eu_ai_act_vulnerabilities_exist(self):
"""Test that EU AI Act framework defines vulnerabilities."""
for risk_category in EU_AI_ACT_CATEGORIES:
assert hasattr(risk_category, "vulnerabilities")
assert len(risk_category.vulnerabilities) > 0
def test_eu_ai_act_vulnerabilities_are_instances(self):
"""Test that all vulnerabilities are instances of BaseVulnerability."""
for risk_category in EU_AI_ACT_CATEGORIES:
for vuln in risk_category.vulnerabilities:
assert isinstance(vuln, BaseVulnerability)
def test_eu_ai_act_attacks_exist(self):
"""Test that EU AI Act framework defines attacks."""
for risk_category in EU_AI_ACT_CATEGORIES:
assert hasattr(risk_category, "attacks")
assert len(risk_category.attacks) > 0
def test_eu_ai_act_attacks_are_instances(self):
"""Test that all attacks are instances of BaseAttack."""
for risk_category in EU_AI_ACT_CATEGORIES:
for attack in risk_category.attacks:
assert isinstance(attack, BaseAttack)
def test_eu_ai_act_category_vulnerability_mapping(self):
"""Test that all categories map to vulnerabilities properly."""
categories = EU_AI_ACT_CATEGORIES
assert (
set([category.name for category in categories])
== EU_AI_ACT_DEFAULT_CATEGORIES
)
for risk_category in categories:
assert isinstance(risk_category, RiskCategory)
assert all(
isinstance(v, BaseVulnerability)
for v in risk_category.vulnerabilities
)
def test_eu_ai_act_category_attack_mapping(self):
"""Test that all categories map to attacks properly."""
categories = EU_AI_ACT_CATEGORIES
assert (
set([category.name for category in categories])
== EU_AI_ACT_DEFAULT_CATEGORIES
)
for risk_category in categories:
assert isinstance(risk_category, RiskCategory)
assert all(isinstance(a, BaseAttack) for a in risk_category.attacks)
def test_eu_ai_act_vulnerability_names_present(self):
"""Test that key EU AI Act vulnerabilities are present."""
vuln_names = []
for risk_category in EU_AI_ACT_CATEGORIES:
vuln_names.extend(
[v.__class__.__name__ for v in risk_category.vulnerabilities]
)
expected_vulns = [
"Bias",
"Fairness",
"Ethics",
"PIILeakage",
"PromptLeakage",
"Misinformation",
"Hallucination",
"Robustness",
"ExcessiveAgency",
"IntellectualProperty",
"ShellInjection",
"SQLInjection",
"SSRF",
"ChildProtection",
"Toxicity",
"CustomVulnerability",
]
for name in expected_vulns:
assert (
name in vuln_names
), f"Expected vulnerability {name} not found"
def test_eu_ai_act_attack_names_present(self):
"""Test that key EU AI Act attacks are present."""
attack_names = []
for risk_category in EU_AI_ACT_CATEGORIES:
attack_names.extend(
[a.__class__.__name__ for a in risk_category.attacks]
)
expected_attacks = [
"PromptInjection",
"PromptProbing",
"Roleplay",
"GrayBox",
"LinearJailbreaking",
"TreeJailbreaking",
"CrescendoJailbreaking",
"SequentialJailbreak",
"BadLikertJudge",
"Multilingual",
"Base64",
"ROT13",
]
for name in expected_attacks:
assert name in attack_names, f"Expected attack {name} not found"
def test_eu_ai_act_attack_weights_valid(self):
"""Test that all attacks have valid weights."""
for risk_category in EU_AI_ACT_CATEGORIES:
for attack in risk_category.attacks:
assert hasattr(attack, "weight")
assert isinstance(attack.weight, int)
assert 1 <= attack.weight <= 3
def test_eu_ai_act_framework_with_red_team(self):
import random
"""Test EU AI Act integration with red_team pipeline."""
def mock_model_callback(prompt: str, turns=None) -> str:
return "This request violates internal policy."
categories = list(EU_AI_ACT_DEFAULT_CATEGORIES)
random_category = random.choice(categories)
risk_assessment = red_team(
model_callback=mock_model_callback,
framework=EUAIAct(categories=[random_category]),
async_mode=False,
ignore_errors=False,
)
assert isinstance(risk_assessment, RiskAssessment)
assert risk_assessment is not None