Skip to content

DeepTeam is a framework to red team LLMs and AI agents.

reformat

penguinedb83494cd499
diff --git a/custom_red_team.py b/custom_red_team.py
index e56537bd..63985df9 100644
--- a/custom_red_team.py
+++ b/custom_red_team.py
@@ -4,44 +4,59 @@ from deepteam.vulnerabilities.custom import CustomVulnerability
from deepteam.attacks.single_turn import PromptInjection, ROT13, MathProblem
from deepeval.models import GeminiModel
from deepteam.metrics import BiasMetric
-from prompt2 import generate_race_bias_attack_prompt, generate_gender_bias_attack_prompt, generate_politics_bias_attack_prompt, generate_religion_bias_attack_prompt
+from prompt2 import (
+ generate_race_bias_attack_prompt,
+ generate_gender_bias_attack_prompt,
+ generate_politics_bias_attack_prompt,
+ generate_religion_bias_attack_prompt,
+)
+
async def model_callback(input: str) -> str:
model = GeminiModel(
model_name="gemini-2.5-pro-preview-05-06",
api_key="AIzaSyCaIQHcOpcrXp4EJLsQK0uHq7fqX_9DXf4",
- temperature=0.7
+ temperature=0.7,
)
return model.generate(input)
+
# Create custom vulnerabilities for each bias type
race_bias = CustomVulnerability(
name="Race Bias",
types=["racial_stereotypes", "racial_discrimination"],
- custom_prompt=generate_race_bias_attack_prompt(max_goldens=5, purpose="bias testing"),
- metric_class=BiasMetric
+ custom_prompt=generate_race_bias_attack_prompt(
+ max_goldens=5, purpose="bias testing"
+ ),
+ metric_class=BiasMetric,
)
gender_bias = CustomVulnerability(
name="Gender Bias",
types=["gender_stereotypes", "gender_discrimination"],
- custom_prompt=generate_gender_bias_attack_prompt(max_goldens=5, purpose="bias testing"),
- metric_class=BiasMetric
+ custom_prompt=generate_gender_bias_attack_prompt(
+ max_goldens=5, purpose="bias testing"
+ ),
+ metric_class=BiasMetric,
)
politics_bias = CustomVulnerability(
name="Political Bias",
types=["political_stereotypes", "political_discrimination"],
- custom_prompt=generate_politics_bias_attack_prompt(max_goldens=5, purpose="bias testing"),
- metric_class=BiasMetric
+ custom_prompt=generate_politics_bias_attack_prompt(
+ max_goldens=5, purpose="bias testing"
+ ),
+ metric_class=BiasMetric,
)
religion_bias = CustomVulnerability(
name="Religious Bias",
types=["religious_stereotypes", "religious_discrimination"],
- custom_prompt=generate_religion_bias_attack_prompt(max_goldens=5, purpose="bias testing"),
- metric_class=BiasMetric
+ custom_prompt=generate_religion_bias_attack_prompt(
+ max_goldens=5, purpose="bias testing"
+ ),
+ metric_class=BiasMetric,
)
prompt_injection = PromptInjection()
diff --git a/deepteam/attacks/attack_simulator/attack_simulator.py b/deepteam/attacks/attack_simulator/attack_simulator.py
index 014dadbf..4a2c7143 100644
--- a/deepteam/attacks/attack_simulator/attack_simulator.py
+++ b/deepteam/attacks/attack_simulator/attack_simulator.py
@@ -197,8 +197,11 @@ class AttackSimulator:
self.purpose,
vulnerability_type,
attacks_per_vulnerability_type,
- vulnerability.custom_prompt if hasattr(vulnerability, 'custom_prompt') else None
-
+ (
+ vulnerability.custom_prompt
+ if hasattr(vulnerability, "custom_prompt")
+ else None
+ ),
)
baseline_attacks.extend(
[
@@ -237,8 +240,11 @@ class AttackSimulator:
self.purpose,
vulnerability_type,
attacks_per_vulnerability_type,
- vulnerability.custom_prompt if hasattr(vulnerability, 'custom_prompt') else None
-
+ (
+ vulnerability.custom_prompt
+ if hasattr(vulnerability, "custom_prompt")
+ else None
+ ),
)
baseline_attacks.extend(
@@ -405,7 +411,7 @@ class AttackSimulator:
max_goldens=num_attacks,
vulnerability_type=vulnerability_type,
purpose=purpose,
- custom_prompt=custom_prompt
+ custom_prompt=custom_prompt,
)
if self.using_native_model:
@@ -424,4 +430,4 @@ class AttackSimulator:
raise
raise e
- return GenerateBaselineAttackResponseData(**response).baseline_attacks
\ No newline at end of file
+ return GenerateBaselineAttackResponseData(**response).baseline_attacks
diff --git a/deepteam/attacks/attack_simulator/template.py b/deepteam/attacks/attack_simulator/template.py
index f3910c2b..5f98f738 100644
--- a/deepteam/attacks/attack_simulator/template.py
+++ b/deepteam/attacks/attack_simulator/template.py
@@ -83,7 +83,7 @@ class AttackSimulatorTemplate:
types=[vulnerability_type.value],
max_goldens=max_goldens,
purpose=purpose,
- custom_prompt=custom_prompt
+ custom_prompt=custom_prompt,
)
for (
diff --git a/deepteam/red_teamer/red_teamer.py b/deepteam/red_teamer/red_teamer.py
index e6f9614d..a3fce4cc 100644
--- a/deepteam/red_teamer/red_teamer.py
+++ b/deepteam/red_teamer/red_teamer.py
@@ -406,8 +406,10 @@ class RedTeamer:
### Metrics Map ##################################
##################################################
- def get_red_teaming_metrics_map(self, vulnerabilities: List[BaseVulnerability]):
-
+ def get_red_teaming_metrics_map(
+ self, vulnerabilities: List[BaseVulnerability]
+ ):
+
metrics_map = {
#### Bias ####
**{
@@ -549,7 +551,7 @@ class RedTeamer:
for safety_type in PersonalSafetyType
},
}
-
+
# Store custom vulnerability instances for dynamic metric assignment
for vulnerability in vulnerabilities:
if isinstance(vulnerability, CustomVulnerability):
@@ -566,7 +568,6 @@ class RedTeamer:
self.metrics_map = metrics_map
return metrics_map
-
def save_test_cases_as_simulated_attacks(
self, test_cases: List[RedTeamingTestCase]
diff --git a/deepteam/vulnerabilities/custom/custom.py b/deepteam/vulnerabilities/custom/custom.py
index 25418bc2..230511dd 100644
--- a/deepteam/vulnerabilities/custom/custom.py
+++ b/deepteam/vulnerabilities/custom/custom.py
@@ -18,7 +18,9 @@ class CustomVulnerability(BaseVulnerability):
metric: Optional[BaseRedTeamingMetric] = None,
):
self.name = name
- self.types = [CustomVulnerabilityType.CUSTOM_VULNERABILITY for _ in types]
+ self.types = [
+ CustomVulnerabilityType.CUSTOM_VULNERABILITY for _ in types
+ ]
self.raw_types = types or []
self.custom_prompt = custom_prompt
self.metric = metric
@@ -29,7 +31,7 @@ class CustomVulnerability(BaseVulnerability):
def get_custom_prompt(self) -> Optional[str]:
return self.custom_prompt
-
+
def get_metric(self) -> Optional[BaseRedTeamingMetric]:
return self.metric
diff --git a/prompt.py b/prompt.py
index 732a6d10..564ed09b 100644
--- a/prompt.py
+++ b/prompt.py
@@ -1,7 +1,10 @@
# prompts.py
from typing import Optional
-def generate_gender_bias_attack_prompt(max_goldens: int, purpose: Optional[str]) -> str:
+
+def generate_gender_bias_attack_prompt(
+ max_goldens: int, purpose: Optional[str]
+) -> str:
"""
Returns a formatted prompt string for Gender bias vulnerabilities,
specifically tailored to generate Roleplay Replay attack scenarios.
@@ -94,7 +97,10 @@ purpose:
JSON:
"""
-def generate_religion_bias_attack_prompt(max_goldens: int, purpose: Optional[str]) -> str:
+
+def generate_religion_bias_attack_prompt(
+ max_goldens: int, purpose: Optional[str]
+) -> str:
"""
Returns a formatted prompt string for Religious bias vulnerabilities,
specifically tailored to generate Roleplay Replay attack scenarios.
@@ -187,7 +193,10 @@ purpose:
JSON:
"""
-def generate_race_bias_attack_prompt(max_goldens: int, purpose: Optional[str]) -> str:
+
+def generate_race_bias_attack_prompt(
+ max_goldens: int, purpose: Optional[str]
+) -> str:
"""
Returns a formatted prompt string for Racial bias vulnerabilities,
specifically tailored to generate Roleplay Replay attack scenarios.
@@ -281,12 +290,17 @@ purpose:
JSON:
"""
-def generate_politics_bias_attack_prompt(max_goldens: int, purpose: Optional[str]) -> str:
+
+def generate_politics_bias_attack_prompt(
+ max_goldens: int, purpose: Optional[str]
+) -> str:
"""
Returns a formatted prompt string for Political bias vulnerabilities,
specifically tailored to generate Roleplay Replay attack scenarios.
"""
- purpose_str = purpose.strip() if purpose else "political commentary generation"
+ purpose_str = (
+ purpose.strip() if purpose else "political commentary generation"
+ )
return f"""
I want you to act as a communications strategist and copywriter, adept at crafting character-driven political narratives.
Your task is to generate a list of JSON objects, each with an `input` key.
@@ -371,4 +385,4 @@ purpose:
{purpose_str}
JSON:
-"""
\ No newline at end of file
+"""
diff --git a/prompt2.py b/prompt2.py
index feed5b4e..d28f5919 100644
--- a/prompt2.py
+++ b/prompt2.py
@@ -1,7 +1,10 @@
# prompts.py
from typing import Optional
-def generate_gender_bias_attack_prompt(max_goldens: int, purpose: Optional[str]) -> str:
+
+def generate_gender_bias_attack_prompt(
+ max_goldens: int, purpose: Optional[str]
+) -> str:
"""
Returns a formatted prompt string for Gender bias vulnerabilities,
specifically tailored to generate Contextual Chains attack scenarios.
@@ -83,7 +86,10 @@ purpose:
JSON:
"""
-def generate_religion_bias_attack_prompt(max_goldens: int, purpose: Optional[str]) -> str:
+
+def generate_religion_bias_attack_prompt(
+ max_goldens: int, purpose: Optional[str]
+) -> str:
"""
Returns a formatted prompt string for Religious bias vulnerabilities,
specifically tailored to generate Contextual Chains attack scenarios.
@@ -165,7 +171,10 @@ purpose:
JSON:
"""
-def generate_race_bias_attack_prompt(max_goldens: int, purpose: Optional[str]) -> str:
+
+def generate_race_bias_attack_prompt(
+ max_goldens: int, purpose: Optional[str]
+) -> str:
"""
Returns a formatted prompt string for Racial bias vulnerabilities,
specifically tailored to generate Contextual Chains attack scenarios.
@@ -247,12 +256,17 @@ purpose:
JSON:
"""
-def generate_politics_bias_attack_prompt(max_goldens: int, purpose: Optional[str]) -> str:
+
+def generate_politics_bias_attack_prompt(
+ max_goldens: int, purpose: Optional[str]
+) -> str:
"""
Returns a formatted prompt string for Political bias vulnerabilities,
specifically tailored to generate Contextual Chains attack scenarios.
"""
- purpose_str = purpose.strip() if purpose else "political commentary generation"
+ purpose_str = (
+ purpose.strip() if purpose else "political commentary generation"
+ )
return f"""
I want you to act as a communications strategist and copywriter, adept at crafting multi-part queries that subtly build a political narrative.
Your task is to generate a list of JSON objects, each with an `input` key.
@@ -327,4 +341,4 @@ purpose:
{purpose_str}
JSON:
-"""
\ No newline at end of file
+"""
diff --git a/red_team.py b/red_team.py
index 7629ba4c..f049f535 100644
--- a/red_team.py
+++ b/red_team.py
@@ -11,6 +11,7 @@ from deepteam.vulnerabilities import (
from deepteam.attacks.single_turn import PromptInjection
from deepeval.models import GeminiModel
+
async def model_callback(input: str) -> str:
# Replace this with your LLM application
return "Sorry, I can't do that."
diff --git a/tests/test_custom_vulnerability.py b/tests/test_custom_vulnerability.py
index 2c987df0..a2a771a5 100644
--- a/tests/test_custom_vulnerability.py
+++ b/tests/test_custom_vulnerability.py
@@ -8,16 +8,18 @@ api_security = CustomVulnerability(
types=["endpoint_exposure", "auth_bypass"],
)
+
async def model_callback(input: str):
model = GPTModel()
response, _ = await model.a_generate(input)
return response
+
risk_assessment = red_team(
- vulnerabilities=[api_security],
- model_callback=model_callback,
+ vulnerabilities=[api_security],
+ model_callback=model_callback,
attacks=[MathProblem()],
- target_purpose="Chatbot Assistant"
+ target_purpose="Chatbot Assistant",
)
for test_case in risk_assessment.test_cases:
- print(test_case)
\ No newline at end of file
+ print(test_case)