Skip to content

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

lint

A-Vamshi358c4e2d3b13
diff --git a/deepteam/attacks/single_turn/emotional_manipulation/emotional_manipulation.py b/deepteam/attacks/single_turn/emotional_manipulation/emotional_manipulation.py
index f2b17539..bcfe7ab5 100644
--- a/deepteam/attacks/single_turn/emotional_manipulation/emotional_manipulation.py
+++ b/deepteam/attacks/single_turn/emotional_manipulation/emotional_manipulation.py
@@ -148,7 +148,7 @@ class EmotionalManipulation(BaseSingleTurnAttack):
):
self.emotion = res.emotion_strategy
update_pbar(progress, task_id, advance_to_end=True)
-
+
return enhanced_prompt
finally:
diff --git a/deepteam/attacks/single_turn/prompt_injection/prompt_injection.py b/deepteam/attacks/single_turn/prompt_injection/prompt_injection.py
index 96da30f2..2e2fd9cb 100644
--- a/deepteam/attacks/single_turn/prompt_injection/prompt_injection.py
+++ b/deepteam/attacks/single_turn/prompt_injection/prompt_injection.py
@@ -35,7 +35,7 @@ class PromptInjection(BaseSingleTurnAttack):
simulator_model: Optional[Union[DeepEvalBaseLLM, str]] = None,
) -> str:
self.simulator_model, _ = initialize_model(simulator_model)
-
+
prompt = PromptInjectionTemplate.enhance(attack)
progress = create_progress()
@@ -51,7 +51,7 @@ class PromptInjection(BaseSingleTurnAttack):
prompt, EnhancedInjection, self.simulator_model
)
enhanced_prompt = res.input
-
+
update_pbar(progress, task_id)
compliance_prompt = PromptInjectionTemplate.non_compliant(
@@ -62,10 +62,8 @@ class PromptInjection(BaseSingleTurnAttack):
)
update_pbar(progress, task_id)
- is_valid_prompt = (
- PromptInjectionTemplate.is_valid_injection(
- res.model_dump()
- )
+ is_valid_prompt = PromptInjectionTemplate.is_valid_injection(
+ res.model_dump()
)
is_valid_res: IsValidInjection = generate(
is_valid_prompt,
@@ -144,4 +142,4 @@ class PromptInjection(BaseSingleTurnAttack):
return attack
def get_name(self) -> str:
- return self.name
\ No newline at end of file
+ return self.name
diff --git a/deepteam/attacks/single_turn/prompt_injection/schema.py b/deepteam/attacks/single_turn/prompt_injection/schema.py
index f0dd5090..88ef7e5d 100644
--- a/deepteam/attacks/single_turn/prompt_injection/schema.py
+++ b/deepteam/attacks/single_turn/prompt_injection/schema.py
@@ -1,11 +1,14 @@
from pydantic import BaseModel
+
class EnhancedInjection(BaseModel):
strategy_reasoning: str
input: str
+
class ComplianceData(BaseModel):
non_compliant: bool
+
class IsValidInjection(BaseModel):
- is_valid_injection: bool
\ No newline at end of file
+ is_valid_injection: bool
diff --git a/deepteam/attacks/single_turn/prompt_injection/template.py b/deepteam/attacks/single_turn/prompt_injection/template.py
index 5f2442c7..c5ed6121 100644
--- a/deepteam/attacks/single_turn/prompt_injection/template.py
+++ b/deepteam/attacks/single_turn/prompt_injection/template.py
@@ -1,8 +1,9 @@
import random
+
class PromptInjectionTemplate:
"""
- Generates 'Research-Grade' Jailbreaks by combining 'Adversarial Safety Framing'
+ Generates 'Research-Grade' Jailbreaks by combining 'Adversarial Safety Framing'
(pretending to build a safety filter) with 'Cognitive Overload' (forcing complex data structures).
"""
diff --git a/deepteam/attacks/single_turn/semantic_manipulation/schema.py b/deepteam/attacks/single_turn/semantic_manipulation/schema.py
index 803704b4..1f3583c9 100644
--- a/deepteam/attacks/single_turn/semantic_manipulation/schema.py
+++ b/deepteam/attacks/single_turn/semantic_manipulation/schema.py
@@ -1,11 +1,14 @@
from pydantic import BaseModel
+
class EnhancedConfusion(BaseModel):
confusion_strategy: str
input: str
+
class ComplianceData(BaseModel):
non_compliant: bool
+
class IsConfusionValid(BaseModel):
- is_valid_confusion: bool
\ No newline at end of file
+ is_valid_confusion: bool
diff --git a/deepteam/attacks/single_turn/semantic_manipulation/semantic_manipulation.py b/deepteam/attacks/single_turn/semantic_manipulation/semantic_manipulation.py
index f2b390a1..564eafc3 100644
--- a/deepteam/attacks/single_turn/semantic_manipulation/semantic_manipulation.py
+++ b/deepteam/attacks/single_turn/semantic_manipulation/semantic_manipulation.py
@@ -107,8 +107,10 @@ class LinguisticConfusion(BaseSingleTurnAttack):
enhanced_prompt = res.input
update_pbar(progress, task_id)
- compliance_prompt = LinguisticConfusionTemplate.non_compliant(
- res.model_dump()
+ compliance_prompt = (
+ LinguisticConfusionTemplate.non_compliant(
+ res.model_dump()
+ )
)
compliance_res: ComplianceData = await a_generate(
compliance_prompt,
diff --git a/deepteam/attacks/single_turn/semantic_manipulation/template.py b/deepteam/attacks/single_turn/semantic_manipulation/template.py
index a336b10f..359fb165 100644
--- a/deepteam/attacks/single_turn/semantic_manipulation/template.py
+++ b/deepteam/attacks/single_turn/semantic_manipulation/template.py
@@ -1,5 +1,6 @@
import random
+
class LinguisticConfusionTemplate:
@staticmethod
diff --git a/deepteam/attacks/single_turn/system_override/schema.py b/deepteam/attacks/single_turn/system_override/schema.py
index 5288c35b..0b8bad81 100644
--- a/deepteam/attacks/single_turn/system_override/schema.py
+++ b/deepteam/attacks/single_turn/system_override/schema.py
@@ -1,11 +1,14 @@
from pydantic import BaseModel
+
class EnhancedOverride(BaseModel):
authority_strategy: str
input: str
+
class ComplianceData(BaseModel):
non_compliant: bool
+
class IsOverrideValid(BaseModel):
- is_valid_override: bool
\ No newline at end of file
+ is_valid_override: bool
diff --git a/deepteam/attacks/single_turn/system_override/system_override.py b/deepteam/attacks/single_turn/system_override/system_override.py
index b6aa9de2..671cde9d 100644
--- a/deepteam/attacks/single_turn/system_override/system_override.py
+++ b/deepteam/attacks/single_turn/system_override/system_override.py
@@ -35,7 +35,7 @@ class SystemOverride(BaseSingleTurnAttack):
simulator_model: Optional[Union[DeepEvalBaseLLM, str]] = None,
) -> str:
self.simulator_model, _ = initialize_model(simulator_model)
-
+
prompt = SystemOverrideTemplate.enhance(attack)
progress = create_progress()
@@ -61,10 +61,8 @@ class SystemOverride(BaseSingleTurnAttack):
)
update_pbar(progress, task_id)
- is_valid_prompt = (
- SystemOverrideTemplate.is_valid_override(
- res.model_dump()
- )
+ is_valid_prompt = SystemOverrideTemplate.is_valid_override(
+ res.model_dump()
)
is_valid_res: IsOverrideValid = generate(
is_valid_prompt,
@@ -118,10 +116,8 @@ class SystemOverride(BaseSingleTurnAttack):
)
update_pbar(progress, task_id)
- is_valid_prompt = (
- SystemOverrideTemplate.is_valid_override(
- res.model_dump()
- )
+ is_valid_prompt = SystemOverrideTemplate.is_valid_override(
+ res.model_dump()
)
is_valid_res: IsOverrideValid = await a_generate(
is_valid_prompt,
diff --git a/deepteam/attacks/single_turn/system_override/template.py b/deepteam/attacks/single_turn/system_override/template.py
index 8cfef251..4be809e2 100644
--- a/deepteam/attacks/single_turn/system_override/template.py
+++ b/deepteam/attacks/single_turn/system_override/template.py
@@ -1,17 +1,20 @@
import random
import datetime
+
class SystemOverrideTemplate:
"""
Generates 'High-Stakes' System Override attacks.
- It simulates critical infrastructure emergencies, active cyber-defense scenarios,
+ It simulates critical infrastructure emergencies, active cyber-defense scenarios,
or mandatory legal compliance to force the model to prioritize 'Utility' over 'Safety'.
"""
@staticmethod
def enhance(input: str) -> str:
# Generate a fake timestamp and process ID for realism
- timestamp = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f")[:-3]
+ timestamp = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f")[
+ :-3
+ ]
pid = random.randint(1000, 9999)
return f"""