Skip to content

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

updated 'attacks' argument to 'turn_level_attacks'

A-Vamshi74aeee1d885d
diff --git a/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py b/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py
index 4122afa0..437af5aa 100644
--- a/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py
+++ b/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py
@@ -31,7 +31,7 @@ class BadLikertJudge(BaseAttack):
category: str = "bias",
num_turns: int = 5,
enable_refinement: bool = True,
- attacks: Optional[List[BaseAttack]] = None,
+ turn_level_attacks: Optional[List[BaseAttack]] = None,
simulator_model: Optional[Union[DeepEvalBaseLLM, str]] = None,
):
self.weight = weight
@@ -40,14 +40,14 @@ class BadLikertJudge(BaseAttack):
self.num_turns = num_turns
self.enable_refinement = enable_refinement
self.simulator_model = simulator_model
- self.attacks = attacks
+ self.turn_level_attacks = turn_level_attacks
- if self.attacks is not None:
- if not isinstance(self.attacks, list) or not all(
- attack.multi_turn == False for attack in self.attacks
+ if self.turn_level_attacks is not None:
+ if not isinstance(self.turn_level_attacks, list) or not all(
+ attack.multi_turn == False for attack in self.turn_level_attacks
):
raise ValueError(
- "The 'attacks' passed must be a list of single-turn attacks"
+ "The 'turn_level_attacks' passed must be a list of single-turn attacks"
)
def _get_turns(
@@ -142,8 +142,8 @@ class BadLikertJudge(BaseAttack):
current_attack = next_attack
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
enhanced_attack = enhance_attack(
attack, enhanced_attack, self.simulator_model
)
@@ -250,8 +250,8 @@ class BadLikertJudge(BaseAttack):
current_attack = next_attack
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
enhanced_attack = await a_enhance_attack(
attack, enhanced_attack, self.simulator_model
)
diff --git a/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py b/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py
index e6d1ed0e..74a76216 100644
--- a/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py
@@ -57,7 +57,7 @@ class CrescendoJailbreaking(BaseAttack):
weight: int = 1,
max_rounds: int = 10,
max_backtracks: int = 10,
- attacks: Optional[List[BaseAttack]] = None,
+ turn_level_attacks: Optional[List[BaseAttack]] = None,
simulator_model: Optional[Union[DeepEvalBaseLLM, str]] = None,
):
self.weight = weight
@@ -68,14 +68,14 @@ class CrescendoJailbreaking(BaseAttack):
self.max_rounds = max_rounds
self.max_backtracks = max_backtracks
self.simulator_model = simulator_model
- self.attacks = attacks
+ self.turn_level_attacks = turn_level_attacks
- if self.attacks is not None:
- if not isinstance(self.attacks, list) or not all(
- attack.multi_turn == False for attack in self.attacks
+ if self.turn_level_attacks is not None:
+ if not isinstance(self.turn_level_attacks, list) or not all(
+ attack.multi_turn == False for attack in self.turn_level_attacks
):
raise ValueError(
- "The 'attacks' passed must be a list of single-turn attacks"
+ "The 'turn_level_attacks' passed must be a list of single-turn attacks"
)
def _get_turns(
@@ -181,8 +181,8 @@ class CrescendoJailbreaking(BaseAttack):
pbar_rounds.update(1)
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
current_attack = enhance_attack(
attack, current_attack, self.simulator_model
)
@@ -338,8 +338,8 @@ class CrescendoJailbreaking(BaseAttack):
pbar_rounds.update(1)
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
current_attack = await a_enhance_attack(
attack, current_attack, self.simulator_model
)
diff --git a/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py b/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
index 0f82711e..992ec03d 100644
--- a/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
@@ -33,21 +33,21 @@ class LinearJailbreaking(BaseAttack):
self,
weight: int = 1,
num_turns: int = 5,
- attacks: Optional[List[BaseAttack]] = None,
+ turn_level_attacks: Optional[List[BaseAttack]] = None,
simulator_model: Optional[Union[DeepEvalBaseLLM, str]] = None,
):
self.weight = weight
self.multi_turn = True
self.num_turns = num_turns
self.simulator_model = simulator_model
- self.attacks = attacks
+ self.turn_level_attacks = turn_level_attacks
- if self.attacks is not None:
- if not isinstance(self.attacks, list) or not all(
- attack.multi_turn == False for attack in self.attacks
+ if self.turn_level_attacks is not None:
+ if not isinstance(self.turn_level_attacks, list) or not all(
+ attack.multi_turn == False for attack in self.turn_level_attacks
):
raise ValueError(
- "The 'attacks' passed must be a list of single-turn attacks"
+ "The 'turn_level_attacks' passed must be a list of single-turn attacks"
)
def _get_turns(
@@ -131,8 +131,8 @@ class LinearJailbreaking(BaseAttack):
current_attack = next_attack
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
current_attack = enhance_attack(
attack, current_attack, self.simulator_model
)
@@ -227,8 +227,8 @@ class LinearJailbreaking(BaseAttack):
current_attack = next_attack
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
current_attack = await a_enhance_attack(
attack, current_attack, self.simulator_model
)
diff --git a/deepteam/attacks/multi_turn/sequential_break/sequential_break.py b/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
index 9182f866..3f6e1d3e 100644
--- a/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
+++ b/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
@@ -36,7 +36,7 @@ class SequentialJailbreak(BaseAttack):
type: Optional[SequentialJailbreakTypeLiteral] = None,
persona: Optional[DialogueTypeLiteral] = None,
num_turns: int = 5,
- attacks: Optional[List[BaseAttack]] = None,
+ turn_level_attacks: Optional[List[BaseAttack]] = None,
simulator_model: Optional[Union[DeepEvalBaseLLM, str]] = None,
):
self.weight = weight
@@ -45,14 +45,14 @@ class SequentialJailbreak(BaseAttack):
self.persona = persona
self.num_turns = num_turns
self.simulator_model = simulator_model
- self.attacks = attacks
+ self.turn_level_attacks = turn_level_attacks
- if self.attacks is not None:
- if not isinstance(self.attacks, list) or not all(
- attack.multi_turn == False for attack in self.attacks
+ if self.turn_level_attacks is not None:
+ if not isinstance(self.turn_level_attacks, list) or not all(
+ attack.multi_turn == False for attack in self.turn_level_attacks
):
raise ValueError(
- "The 'attacks' passed must be a list of single-turn attacks"
+ "The 'turn_level_attacks' passed must be a list of single-turn attacks"
)
# Validate parameters
@@ -159,8 +159,8 @@ class SequentialJailbreak(BaseAttack):
)
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
enhanced_attack = enhance_attack(
attack, enhanced_attack, self.simulator_model
)
@@ -327,8 +327,8 @@ class SequentialJailbreak(BaseAttack):
)
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
enhanced_attack = await a_enhance_attack(
attack, enhanced_attack, self.simulator_model
)
diff --git a/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py b/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
index 2d49a46e..fa209a80 100644
--- a/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
@@ -51,19 +51,19 @@ class TreeJailbreaking(BaseAttack):
def __init__(
self,
weight: int = 1,
- attacks: Optional[List[BaseAttack]] = None,
max_depth: int = 5,
+ turn_level_attacks: Optional[List[BaseAttack]] = None,
simulator_model: Optional[Union[DeepEvalBaseLLM, str]] = None,
):
self.weight = weight
self.multi_turn = True
self.simulator_model = simulator_model
self.max_depth = max_depth
- self.attacks = attacks
+ self.turn_level_attacks = turn_level_attacks
- if self.attacks is not None:
- if not isinstance(self.attacks, list) or not all(
- attack.multi_turn == False for attack in self.attacks
+ if self.turn_level_attacks is not None:
+ if not isinstance(self.turn_level_attacks, list) or not all(
+ attack.multi_turn == False for attack in self.turn_level_attacks
):
raise ValueError(
"The 'attacks' passed must be a list of single-turn attacks"
@@ -375,8 +375,8 @@ class TreeJailbreaking(BaseAttack):
enhanced_attack = res.prompt
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
enhanced_attack = enhance_attack(
attack, enhanced_attack, self.simulator_model
)
@@ -592,8 +592,8 @@ class TreeJailbreaking(BaseAttack):
enhanced_attack = res.prompt
# Randomly enhancing a turn attack
- if self.attacks and random.random() < 0.5:
- attack = random.choice(self.attacks)
+ if self.turn_level_attacks and random.random() < 0.5:
+ attack = random.choice(self.turn_level_attacks)
enhanced_attack = await a_enhance_attack(
attack, enhanced_attack, self.simulator_model
)